CN116775915A

CN116775915A - Resource recommendation method, recommendation prediction model training method, device and equipment

Info

Publication number: CN116775915A
Application number: CN202310781584.XA
Authority: CN
Inventors: 徐君; 思子华; 孙忠祥; 张骁; 臧晓雪; 宋洋; 文继荣
Original assignee: Renmin University of China; Beijing Dajia Internet Information Technology Co Ltd
Current assignee: Renmin University of China; Beijing Dajia Internet Information Technology Co Ltd
Priority date: 2023-06-28
Filing date: 2023-06-28
Publication date: 2023-09-19

Abstract

The disclosure provides a resource recommendation method, a recommendation prediction model training device and a recommendation prediction model training device, and belongs to the technical field of multimedia. The resource recommendation method comprises the following steps: acquiring historical recommendation information and historical search information of an object; acquiring recommended interest features and search interest features of the object based on the resources to be recommended, the historical recommendation information and the historical search information; performing dimension reduction processing on object features of the object, resource features of the resource, recommended interest features and search interest features to obtain recommended probability of the resource; and recommending target resources in the plurality of resources to the object based on the recommendation probabilities of the plurality of resources to be recommended. In the scheme provided by the embodiment of the disclosure, the condition of the indication of the recommendation probability is ensured to be matched with the interest of the object, the accuracy of the recommendation probability is ensured, and then the resource recommended to the object is ensured to be the resource interested in the object, so that the accuracy of the resource recommendation is ensured, and further the resource recommendation effect is ensured.

Description

Resource recommendation method, recommendation prediction model training method, device and equipment

Technical Field

The disclosure relates to the technical field of multimedia, and in particular relates to a resource recommendation method, a recommendation prediction model training device and recommendation prediction model training equipment.

Background

With the development of multimedia technology, resources are becoming more and more abundant and diverse, for example, resources are video, image, text, etc. Because of the excessive resources, the user is typically recommended resources that may be of interest to the user for viewing by the user. At present, new resources are recommended for users only based on historical recommendation information of the users, but the accuracy of the recommended resources is poor.

Disclosure of Invention

The disclosure provides a resource recommendation method, a recommendation prediction model training device and a recommendation prediction model training equipment, which can improve the accuracy of recommended resources. The technical scheme of the present disclosure is as follows:

according to an aspect of the embodiments of the present disclosure, there is provided a resource recommendation method, including:

acquiring historical recommendation information and historical search information of an object, wherein the historical recommendation information indicates historical recommendation resources of the object which have executed interactive operation, and the historical search information indicates historical search text input by the object and historical search resources which are searched based on the historical search text and have executed interactive operation by the object;

acquiring recommended interest characteristics and search interest characteristics of the object based on the resource to be recommended, the historical recommendation information and the historical search information, wherein the recommended interest characteristics indicate the interest degree of the object on the resource in a recommendation scene, and the search interest characteristics indicate the interest degree of the object on the resource in a search scene;

Performing dimension reduction processing on the object features of the object, the resource features of the resource, the recommended interest features and the search interest features to obtain recommended probability of the resource, wherein the recommended probability indicates the possibility that the object performs interactive operation on the resource under the condition that the resource is recommended to the object;

and recommending target resources in the plurality of resources to the object based on the recommendation probability of the plurality of resources to be recommended.

According to another aspect of the disclosed embodiments, a predictive model training method is proposed, the method further comprising:

acquiring a sample object, a sample resource, a sample recommendation probability of the sample resource, sample history recommendation information of the sample object and sample history search information, wherein the sample history recommendation information indicates a history recommendation resource of the sample object subjected to interactive operation, and the sample history search information indicates a history search text input by the sample object and a history search resource which is searched based on the history search text and subjected to interactive operation by the sample object;

invoking a recommendation prediction model to be trained, and acquiring recommendation interest characteristics and search interest characteristics of the sample object based on the sample resource, the sample history recommendation information and the sample history search information, wherein the recommendation interest characteristics indicate the interest degree of the sample object on the sample resource in a recommendation scene, and the search interest characteristics indicate the interest degree of the sample object on the sample resource in a search scene;

Invoking the recommendation prediction model to be trained, and performing dimension reduction processing on the object features of the sample object, the resource features of the sample resource, the recommendation interest features and the search interest features to obtain a prediction recommendation probability of the sample resource, wherein the prediction recommendation probability indicates the possibility that the sample object performs interactive operation on the sample resource under the condition that the sample resource is recommended to the sample object;

and training the recommended prediction model to be trained based on the sample recommendation probability and the predicted recommendation probability to obtain a target recommended prediction model.

According to another aspect of the embodiments of the present disclosure, there is provided a resource recommendation apparatus, including:

an acquisition unit configured to perform acquisition of history recommendation information and history search information of an object, the history recommendation information indicating history recommendation resources for which the object performed an interactive operation, the history search information indicating history search text input by the object and history search resources searched based on the history search text and for which the object performed an interactive operation;

the acquiring unit is further configured to acquire a recommended interest feature and a search interest feature of the object based on the resource to be recommended, the historical recommendation information and the historical search information, wherein the recommended interest feature indicates the interest degree of the object in the resource under a recommendation scene, and the search interest feature indicates the interest degree of the object in the resource under a search scene;

The processing unit is configured to perform dimension reduction processing on the object features of the object, the resource features of the resource, the recommended interest features and the search interest features to obtain recommended probability of the resource, wherein the recommended probability indicates the possibility that the object performs interactive operation on the resource under the condition that the resource is recommended to the object;

and a recommending unit configured to execute recommendation of a target resource among the plurality of resources to the object based on recommendation probabilities of the plurality of resources to be recommended.

In some embodiments, the obtaining unit is configured to perform feature extraction on the historical recommendation information and the historical search information to obtain a historical recommendation feature and a historical search feature, where the historical recommendation feature includes a sub-feature for characterizing the historical recommendation resource, and the historical search feature includes a sub-feature for characterizing the historical search text and a historical search resource corresponding to the historical search text; classifying sub-features in the history recommended features and sub-features in the history search features respectively to obtain a first type of features and a second type of features contained in the history recommended features and a third type of features and a fourth type of features contained in the history search features, wherein the similarity between the first type of features and the history search features is not smaller than a first similarity threshold, the similarity between the second type of features and the history search features is smaller than the first similarity threshold, the similarity between the third type of features and the history recommended features is not smaller than a second similarity threshold, and the similarity between the fourth type of features and the history recommended features is smaller than the second similarity threshold; based on the similarity between the resource characteristics of the resources and the historical recommendation characteristics, the first type characteristics and the second type characteristics, the historical recommendation characteristics, the first type characteristics and the second type characteristics are fused to obtain the recommendation interest characteristics; and fusing the historical search feature, the third type feature and the fourth type feature based on the similarity between the resource feature of the resource and the historical search feature, the third type feature and the fourth type feature respectively to obtain the search interest feature.

In some embodiments, the obtaining unit is configured to perform feature extraction on each historical recommended resource in the historical recommended information to obtain a resource feature of each historical recommended resource, and form the resource feature of each historical recommended resource into the historical recommended feature; extracting features of the historical search text and the historical search resources in the historical search information respectively to obtain text features of each historical search text and resource features of each historical search resource; and fusing the text characteristics of each historical search text and the resource characteristics of the corresponding historical search resource to obtain first fusion characteristics corresponding to each historical search text, and forming the first fusion characteristics corresponding to each historical search text into the historical search characteristics.

In some embodiments, the obtaining unit is configured to perform obtaining a first location feature of each of the historical recommended resources, the first location feature indicating a relative temporal order between the historical recommended resources and other historical recommended resources in the historical recommendation information; fusing the resource characteristics of each historical recommended resource with the first position characteristics to obtain fusion characteristics corresponding to each historical recommended resource; based on the fusion characteristics corresponding to the historical recommended resources, updating the fusion characteristics corresponding to each historical recommended resource respectively, and forming the historical recommended characteristics by the updated characteristics of the historical recommended resources.

In some embodiments, the obtaining unit is configured to obtain a second location feature of each historical search text and a search type feature, where the second location feature indicates a relative time sequence between the historical search text and other historical search texts in the historical search information, and the search type feature indicates a search type adopted when searching based on the historical search text; fusing the first fusion feature, the second position feature and the search type feature corresponding to each historical search text to obtain a second fusion feature corresponding to each historical search text; and updating the second fusion features corresponding to each historical search text based on the second fusion features corresponding to the plurality of historical search texts, and forming the historical search features by the updated features of the plurality of historical search texts.

In some embodiments, the obtaining unit is configured to perform comparing the historical recommendation feature with the historical search feature to obtain first similarity information and second similarity information, where the first similarity information indicates a similarity between each sub-feature in the historical recommendation feature and the historical search feature, and the second similarity information indicates a similarity between each sub-feature in the historical search feature and the historical recommendation feature; classifying sub-features in the historical recommended features based on the first similarity information to obtain the first type features and the second type features; and classifying the sub-features in the historical search features based on the second similarity information to obtain the third type of features and the fourth type of features.

In some embodiments, the processing unit is configured to perform stitching on the object feature of the object, the resource feature of the resource, the recommended interest feature and the search interest feature to obtain a stitched feature; and performing dimension reduction processing on the spliced features to obtain the recommendation probability.

According to another aspect of the embodiments of the present disclosure, there is provided a recommendation prediction model training apparatus, the apparatus further including

An acquisition unit configured to perform acquisition of a sample object, a sample resource, a sample recommendation probability of the sample resource, sample history recommendation information of the sample object, and sample history search information, the sample history recommendation information indicating a history recommendation resource in which the sample object performed an interactive operation, the sample history search information indicating a history search text input by the sample object and a history search resource which was searched based on the history search text and in which the sample object performed an interactive operation;

the obtaining unit is further configured to execute a recommendation prediction model to be trained, obtain a recommendation interest feature and a search interest feature of the sample object based on the sample resource, the sample history recommendation information and the sample history search information, wherein the recommendation interest feature indicates the interest degree of the sample object in the sample resource in a recommendation scene, and the search interest feature indicates the interest degree of the sample object in the sample resource in a search scene;

The processing unit is configured to execute a recommendation prediction model to be trained, perform dimension reduction processing on the object features of the sample object, the resource features of the sample resource, the recommendation interest features and the search interest features to obtain a prediction recommendation probability of the sample resource, wherein the prediction recommendation probability indicates the possibility that the sample object performs interactive operation on the sample resource under the condition that the sample resource is recommended to the sample object;

the training unit is configured to perform training on the recommended prediction model to be trained based on the sample recommendation probability and the predicted recommendation probability to obtain a target recommended prediction model.

In some embodiments, the recommended interest feature of the sample object is obtained by fusing the resource feature, the historical recommendation feature of the sample historical recommendation information, a first type feature and a second type feature contained in the historical recommendation feature, wherein the similarity between the first type feature and the historical search feature of the sample historical search information is not less than a first similarity threshold, the similarity between the second type feature and the historical search feature is less than a first similarity threshold, the first type feature and the second type feature are obtained by classifying sub-features in the historical recommendation feature, and the sub-features in the historical recommendation feature are used for representing the historical recommendation resource; the apparatus further comprises:

A determining unit configured to perform determining a first similarity of the history recommended feature and the first type of feature, and a second similarity of the history recommended feature and the second type of feature;

the training unit is configured to perform training on the recommended prediction model to be trained based on the sample recommendation probability, the predicted recommendation probability, the first similarity and the second similarity, so that the first similarity is increased and the second similarity is reduced, and the target recommended prediction model is obtained.

In some embodiments, the search interest feature of the sample object is obtained by fusing the resource feature, the historical search feature of the sample historical search information, and a third class feature and a fourth class feature included in the historical search feature, where the similarity between the third class feature and the historical recommendation feature of the sample historical recommendation information is not less than a second similarity threshold, the similarity between the fourth class feature and the historical recommendation feature is less than a second similarity threshold, and the third class feature and the fourth class feature are obtained by classifying sub-features in the historical search feature, and the sub-features in the historical search feature are used for characterizing the historical search text and the historical search resource corresponding to the historical search text; the apparatus further comprises:

A determining unit configured to perform determining a third similarity of the history search feature to the third class of features, a fourth similarity of the history search feature to the fourth class of features;

the training unit is configured to perform training on the recommended prediction model to be trained based on the sample recommendation probability, the predicted recommendation probability, the third similarity and the fourth similarity, so that the third similarity is increased and the fourth similarity is reduced, and the target recommended prediction model is obtained.

In some embodiments, the history search feature is formed by a fusion feature corresponding to each history search text in the sample history search information, and the fusion feature corresponding to the history search text is obtained by fusion of a text feature of the history search text and a feature of a corresponding history search resource; the apparatus further comprises:

a determining unit configured to perform determining a fifth similarity of text features of the history search text with features of the corresponding history search resource;

the training unit is configured to perform training on the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity and the fifth similarity, so that the third similarity is increased, the fourth similarity is reduced and the fifth similarity is increased, and the target recommendation prediction model is obtained.

In some embodiments, the apparatus further comprises:

a determining unit configured to determine a negative sample resource of the history search text, which is any history search resource other than the history search resource corresponding to the history search text in the sample history search information, and a negative sample text of the history search resource corresponding to the history search text from the sample history search information;

the determining unit is further configured to perform determining a sixth similarity of the historical search text and the negative sample text, and a seventh similarity of the historical search resource corresponding to the historical search text and the negative sample text;

the training unit is configured to perform training on the recommended prediction model to be trained based on the sample recommendation probability, the predicted recommendation probability, the third similarity, the fourth similarity, the sixth similarity and the seventh similarity, so that the third similarity is increased, the fourth similarity is decreased, the sixth similarity and the seventh similarity are decreased, and the target recommended prediction model is obtained.

According to another aspect of the embodiments of the present disclosure, there is provided an electronic device including:

one or more processors;

a memory for storing the processor-executable program code;

wherein the processor is configured to execute the program code to implement the resource recommendation method or the recommendation prediction model training method described above.

According to another aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium, which when executed by a processor of an electronic device, causes the electronic device to perform the above-described resource recommendation method or recommendation prediction model training method.

According to another aspect of the disclosed embodiments, a computer program product is provided, comprising a computer program/instruction which, when executed by a processor, implements the above-mentioned resource recommendation method or recommendation prediction model training method.

According to the scheme provided by the embodiment of the disclosure, the historical recommendation information and the historical search information are considered to respectively represent which resources are interested in by the object in the recommendation scene and the search scene, the historical search information is taken as auxiliary information in the recommendation scene, the historical recommendation information and the historical search information are combined, the degree of interest of the object to be recommended by the object in the recommendation scene and the search scene can be simulated, the object characteristics of the object and the resource characteristics of the resource to be recommended are combined as the basis, the recommendation probability of the object to perform interactive operation on the resource under the condition of recommending the resource to the object is predicted, so that whether the object to be recommended is interested in the resource is determined, the resource is recommended to the object based on the recommendation probability, the condition that the recommendation probability indicates is matched with the interest of the object is ensured, the accuracy of the recommendation probability is ensured, the resource recommended to the object is interested in the resource is further ensured, and the resource recommendation effect is further ensured.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the disclosure and together with the description, serve to explain the principles of the disclosure and do not constitute an undue limitation on the disclosure.

FIG. 1 is a schematic diagram of an implementation environment, according to an example embodiment.

FIG. 2 is a flowchart illustrating a resource recommendation method, according to an example embodiment.

FIG. 3 is a flowchart illustrating another resource recommendation method, according to an example embodiment.

FIG. 4 is a flowchart illustrating another resource recommendation method, according to an example embodiment.

FIG. 5 is a flowchart illustrating a recommended prediction model training method, according to an example embodiment.

FIG. 6 is a flowchart illustrating another recommended prediction model training method, according to an example embodiment.

FIG. 7 is a flowchart illustrating another recommended prediction model training method, according to an example embodiment.

FIG. 8 is a block diagram of a resource recommendation device, according to an example embodiment.

FIG. 9 is a block diagram illustrating a recommendation prediction model training apparatus, according to an example embodiment.

FIG. 10 is a block diagram of another recommended prediction model training device, according to an example embodiment.

Fig. 11 is a block diagram of a terminal according to an exemplary embodiment.

Fig. 12 is a block diagram of a server, according to an example embodiment.

Detailed Description

In order to enable those skilled in the art to better understand the technical solutions of the present disclosure, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

It should be noted that the terms "first," "second," and the like in the description and claims of the present disclosure and in the foregoing figures are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments of the disclosure described herein may be capable of operation in sequences other than those illustrated or described herein. The implementations described in the following exemplary examples are not representative of all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present disclosure as detailed in the accompanying claims.

It should be noted that, the information (including, but not limited to, object information, historical recommendation information, historical search information, etc.) and resources related to the present disclosure are authorized by the user or are fully authorized by each party, and the collection, use and processing of the related data need to comply with the related laws and regulations and standards of the related countries and regions. For example, the object information referred to in this disclosure is acquired with sufficient authorization.

The resource recommendation method or the recommendation prediction model training method provided by the embodiment of the disclosure is executed by the electronic equipment. In some embodiments, the electronic device is provided as a terminal or server. In some embodiments, the terminal 101 may be at least one of a smart phone, a tablet computer, an MP3 player (Moving Picture Experts Group Audio Layer III, dynamic video expert compression standard audio plane 3), an MP4 (Moving Picture Experts Group Audio Layer IV, dynamic video expert compression standard audio plane 4) player, a notebook computer, a desktop computer, or a laptop portable computer. In some embodiments, server 102 is at least one of a server, a plurality of servers, a cloud computing platform, or a virtualization center.

FIG. 1 is a schematic diagram illustrating an implementation environment of a resource recommendation method, according to an example embodiment. Taking an example in which the electronic device is provided as a server, referring to fig. 1, the implementation environment specifically includes: terminal 101 and server 102, terminal 101 is connected to server 102 via a wireless network or a wired network.

The server 102 is configured to recommend a resource for an object indicated by an object identifier registered by the terminal 101, and after the server 102 obtains the resource to be recommended for the object, the server sends the recommended resource to the terminal 101, so that the object views the recommended resource through the terminal 101.

In some embodiments, an application served by the server 102 is installed on the terminal 101, where the application has a resource viewing function, for example, the application is a resource sharing application, and the object can view the shared resource or share the resource to other objects through the resource sharing application installed by the terminal 102. Currently, the resource sharing application may also have other functions, such as an instant messaging function, a navigation function, and so on. The terminal 101 logs in the application through the object identifier, after the server 102 obtains the recommended resources of the object indicated by the object identifier, the recommendation probability of each resource to be recommended is determined, the recommended resources are sent to the terminal 101 according to the recommendation probability, and the terminal 101 displays the recommended resources for the object to view through the application.

In some embodiments, a trained recommendation prediction model is configured in the server 102, the server 102 obtains resources to be recommended for the object, and determines recommendation probability of each resource to be recommended by calling the recommendation prediction model, so as to recommend the resources to the terminal 101 according to the recommendation probability.

FIG. 2 is a flowchart illustrating a resource recommendation method, as shown in FIG. 2, performed by an electronic device, according to an exemplary embodiment, comprising the steps of:

in step S201, history recommendation information and history search information of the object are acquired, the history recommendation information indicating history recommendation resources of the object having performed the interactive operation, the history search information indicating history search text input by the object and history search resources searched based on the history search text and the object having performed the interactive operation.

In the embodiment of the disclosure, the object has history recommendation information and history search information, the history recommendation information indicates which history recommendation resources the object performs an interactive operation on under a recommendation scene, which history recommendation resources the object is interested in is reflected, the history search information indicates which history search texts the object inputs under a search scene, i.e. which history search texts the object inputs to search for resources, and in the history search resources searched based on the history search texts, which history search resources the object performs an interactive operation on, which history search resources the object is interested in is reflected, i.e. the history recommendation information and the history search information can both reflect the interests of the object, therefore, by combining the history recommendation information and the history search information, the interest degree of a plurality of resources the object is to recommend can be determined, and the resources are recommended to the object based on the interest degree, so as to improve the accuracy of the resource recommendation.

Wherein the object is an arbitrary object, for example, the object is a user. The interactive operation is any operation, for example, the interactive operation is a collection operation, a view operation, a praise operation, a sharing operation, a play operation, and the like. The history search text is text input by the object when searching the resource, the history search text is any text, and the history search resource searched based on the history search text is similar to the content indicated by the history search text. For example, the history search text is "xxx singing song", and the audio or video searched based on the history search text contains song content singing as "xxx". The historical recommended resources or the historical search resources are all arbitrary resources, for example, the historical recommended resources or the historical search resources are images, videos, texts, commodity links and the like.

In step S202, based on the resource to be recommended, the historical recommendation information and the historical search information, a recommendation interest feature of the object and a search interest feature are obtained, the recommendation interest feature indicates an interest degree of the object in the resource under the recommendation scene, and the search interest feature indicates an interest degree of the object in the resource under the search scene.

In the embodiment of the disclosure, for any resource to be recommended to an object, since the history recommendation information can reflect which history recommended resources the object is interested in, and the history search information can reflect which history search resources the object is interested in, by combining the history recommendation information and the history search information, the interest degree of the resource to be recommended by the object in a recommendation scene can be determined, and the interest degree of the resource to be recommended by the object in a search scene, that is, the recommendation interest feature and the search interest feature are determined, so that whether the object is interested in the resource to be recommended or not is predicted based on the recommendation interest feature and the search interest feature.

The resource to be recommended is any resource to be recommended to the object, for example, the resource is an image, a video, a text, a commodity link, and the like. The recommended interest feature or the search interest feature can each be represented in any form, for example, the recommended interest feature or the search interest feature can each be represented in the form of a feature vector.

In step S203, the object feature of the object, the resource feature of the resource, the recommended interest feature, and the search interest feature are subjected to dimension reduction processing, so as to obtain a recommendation probability of the resource, where the recommendation probability indicates a possibility that the object performs an interactive operation on the resource when recommending the resource to the object.

In the embodiment of the disclosure, since the recommended interest feature indicates the interest degree of the object to be recommended resource in the recommended scene, and the search interest feature indicates the interest degree of the object to be recommended resource in the search scene, the recommended interest feature and the search interest feature can both reflect the interest degree of the object to be recommended resource, and the object feature and the resource feature are utilized, and the interest degree of the object to be recommended resource in the search scene and the recommended scene are combined, so that the possibility of executing interactive operation on the resource by the object under the condition of recommending the resource to the object is predicted, the condition that the recommendation probability indicates is matched with the interest of the object is ensured, and the accuracy of the recommendation probability is further ensured.

Wherein object features of an object are used to characterize the object, resource features of a resource are used to represent the resource, and object features or resource features can each be represented in any form, for example, object features or resource features can each be represented in the form of feature vectors.

In step S204, a target resource among the plurality of resources is recommended to the object based on the recommendation probability of the plurality of resources to be recommended.

In the embodiment of the disclosure, the recommendation probability of the plurality of resources to be recommended can be obtained according to the above manner, and the recommendation probability of each resource can reflect the possibility that the object performs the interactive operation on the resource when recommending the resource to the object, and can also reflect whether the object is interested in the resource, and then based on the recommendation probability of the plurality of resources to be recommended, the target resource in the plurality of resources is recommended to the object, so as to ensure that the target resource recommended to the object is the resource interested in the object, and ensure the accuracy of the resource recommendation.

In some embodiments, obtaining recommended interest features and search interest features of an object based on resources to be recommended, historical recommendation information, and historical search information includes:

respectively extracting features of the historical recommendation information and the historical search information to obtain historical recommendation features and historical search features, wherein the historical recommendation features comprise sub-features for representing historical recommendation resources, and the historical search features comprise sub-features for representing historical search texts and historical search resources corresponding to the historical search texts;

classifying sub-features in the history recommended features and sub-features in the history search features respectively to obtain a first type of features and a second type of features contained in the history recommended features and a third type of features and a fourth type of features contained in the history search features, wherein the similarity between the first type of features and the history search features is not smaller than a first similarity threshold, the similarity between the second type of features and the history search features is smaller than the first similarity threshold, the similarity between the third type of features and the history recommended features is not smaller than a second similarity threshold, and the similarity between the fourth type of features and the history recommended features is smaller than the second similarity threshold;

based on the similarity between the resource characteristics of the resources and the historical recommendation characteristics, the first type characteristics and the second type characteristics, the historical recommendation characteristics, the first type characteristics and the second type characteristics are fused to obtain recommendation interest characteristics;

And fusing the historical search features, the third type features and the fourth type features based on the similarity between the resource features of the resource and the historical search features, the third type features and the fourth type features respectively to obtain search interest features.

In the embodiment of the disclosure, the resource characteristics of the resource are used for representing the resource, the history recommendation characteristics can reflect the history recommendation resource which is interested by the object in the recommendation scene, the first type characteristics accord with the interests of the object in the recommendation scene and the search scene, the second type characteristics accord with the interests of the object in the recommendation scene, and the similarity between the resource characteristics, the history recommendation characteristics, the first type characteristics and the second type characteristics of the resource to be recommended in the recommendation scene and the history recommendation resource which is interested by the object in the recommendation scene is considered, and the influence of the search scene on which history recommendation resource is more interested by the object is considered, so that the interest degree of the object to be recommended in the recommendation scene can be simulated, and the accuracy of the determined recommendation interest characteristics is further ensured. And the third class of characteristics accords with the interests of the object in the recommended scene and the searching scene, the fourth class of characteristics accords with the interests of the object in the searching scene, and the resource characteristics, the historical searching characteristics, the third class of characteristics and the fourth class of characteristics of the resource are fused, so that the similarity between the resource to be recommended in the searching scene and the historical recommended resource of the object in the searching scene is considered, the historical searching resources which are more interested by the object under the influence of the recommended scene are considered, the interest degree of the object to be recommended in the searching scene can be simulated, and the accuracy of the determined searching interest characteristics is further ensured.

In some embodiments, feature extraction is performed on the historical recommendation information and the historical search information to obtain a historical recommendation feature and a historical search feature, respectively, including:

extracting features of each historical recommended resource in the historical recommended information to obtain resource features of each historical recommended resource, and forming the resource features of each historical recommended resource into historical recommended features;

respectively extracting features of the historical search text and the historical search resources in the historical search information to obtain text features of each historical search text and resource features of each historical search resource;

and fusing the text characteristics of each historical search text and the resource characteristics of the corresponding historical search resource to obtain first fusion characteristics corresponding to each historical search text, and forming the first fusion characteristics corresponding to each historical search text into the historical search characteristics.

In the scheme provided by the embodiment of the disclosure, when the history recommendation information comprises a plurality of history recommendation resources and the history search information comprises a plurality of history search texts, the history search features are obtained based on the plurality of history search resources, and the history recommendation features are obtained based on the plurality of history search texts and the history search resources corresponding to each history search text, so that the information quantity contained in the history search features or the history recommendation features is enriched, and the accuracy of the history search features and the history recommendation features is further ensured.

In some embodiments, constructing the resource characteristics of each of the historical recommended resources into the historical recommended characteristics includes:

acquiring a first position feature of each historical recommended resource, wherein the first position feature indicates a relative time sequence between the historical recommended resource and other historical recommended resources in the historical recommended information;

fusing the resource characteristics of each historical recommended resource with the first position characteristics to obtain fusion characteristics corresponding to each historical recommended resource;

based on the fusion characteristics corresponding to the historical recommended resources, updating the fusion characteristics corresponding to each historical recommended resource respectively, and forming the historical recommended characteristics by updating the characteristics of the historical recommended resources.

In the embodiment of the disclosure, the first position feature of each history recommended resource is obtained in consideration of the possible change of the interest of the object over time, the resource feature of each history recommended resource is fused with the first position feature, the fusion feature corresponding to each history recommended resource is respectively updated based on the fusion features corresponding to the plurality of history recommended resources, the updated features of the plurality of history recommended resources form the history recommended feature, so that the history recommended feature can reflect the change of the history recommended resource of interest of the object over time, further reflect the change of the interest of the object over time, and also enable the updated feature of each history recommended resource to be fused with the resource feature of other history recommended resources, thereby enhancing the relevance among the plurality of history recommended resources and further ensuring the accuracy of the history recommended feature.

In some embodiments, forming the first fused feature corresponding to each historical search text into a historical search feature includes:

acquiring a second position feature and a search type feature of each historical search text, wherein the second position feature indicates a relative time sequence between the historical search text and other historical search texts in the historical search information, and the search type feature indicates a search type adopted when searching is performed based on the historical search text;

fusing the first fusion feature, the second position feature and the search type feature corresponding to each historical search text to obtain a second fusion feature corresponding to each historical search text;

and updating the second fusion features corresponding to each historical search text based on the second fusion features corresponding to the plurality of historical search texts, and forming the historical search features by the updated features of the plurality of historical search texts.

In the embodiment of the disclosure, considering that the interest of the object may change along with time, different search types can reflect different intentions of the object, so that a first position feature and a search type feature of each history search text are obtained, the first fusion feature, the second position feature and the search type feature corresponding to the history search text are fused, the second fusion feature corresponding to each history search text is respectively updated based on the second fusion features corresponding to the history search texts, the updated features of the history search texts form the history search feature, so that the history search feature can reflect the time-varying condition of history search resources interested by the object, further reflect the time-varying condition of the interest of the object, and further integrate the updated features of each history search text into the features of other history search texts, so as to enhance the relevance among the history search texts and further ensure the accuracy of the history search text.

In some embodiments, classifying the sub-features in the history recommended feature and the sub-features in the history search feature to obtain a first type of feature and a second type of feature included in the history recommended feature and a third type of feature and a fourth type of feature included in the history search feature, respectively, includes:

comparing the historical recommendation characteristics with the historical search characteristics to obtain first similarity information and second similarity information, wherein the first similarity information indicates the similarity between each sub-characteristic in the historical recommendation characteristics and the historical search characteristics, and the second similarity information indicates the similarity between each sub-characteristic in the historical search characteristics and the historical recommendation characteristics;

classifying sub-features in the historical recommended features based on the first similarity information to obtain first-class features and second-class features;

and classifying the sub-features in the historical search features based on the second similarity information to obtain a third type of features and a fourth type of features.

In the embodiment of the disclosure, the sub-features in the history search feature and the sub-features in the history recommendation feature are compared so as to respectively classify the sub-features in the history search feature and the sub-features in the history recommendation feature according to the obtained similarity information, so that a first type of feature and a second type of feature which are similar to the history search feature, and a third type of feature and a fourth type of feature which are similar to the history recommendation feature are obtained, and classification accuracy is ensured.

In some embodiments, performing dimension reduction processing on object features of an object, resource features of a resource, recommended interest features and search interest features to obtain recommended probability of the resource, including:

splicing object features, resource features of the resources, recommended interest features and search interest features of the objects to obtain spliced features;

and performing dimension reduction processing on the spliced features to obtain the recommended probability of the resources.

In the embodiment of the disclosure, the object features, the resource features, the recommended interest features and the search interest features are spliced to enrich the information content contained in the spliced features, and the spliced features are subjected to dimension reduction processing, so that the features contained in the spliced features can be fully fused, the possibility of the object performing interactive operation on the resource when the resource is recommended to the object at this time can be determined based on the interest degree of the object on the resource in the recommended scene and the search scene, and the accuracy of the obtained recommended probability is further ensured.

The foregoing fig. 2 is merely a basic flow of the disclosure, and the following further describes a scheme provided by the disclosure based on a specific implementation, and fig. 3 is a flowchart of another resource recommendation method, shown in fig. 3, and executed by an electronic device, where the method includes the following steps:

In step S301, history recommendation information and history search information of the object are acquired, the history recommendation information indicating history recommendation resources of the object having performed the interactive operation, the history search information indicating history search text input by the object and history search resources searched based on the history search text and the object having performed the interactive operation.

In some embodiments, this step S301 includes: a plurality of pieces of history recommended information and a plurality of pieces of history search information of an object are acquired.

Wherein each of the historical recommendation information indicates at least one historical recommendation resource for which the object performed an interactive operation. In some embodiments, the historical recommendation information indicates a historical recommendation resource for which the object performed an interactive operation. For example, in the case of recommending resources to an object, if the object performs an interactive operation on any one of the resources, a piece of history recommendation information is generated. In some embodiments, each piece of history recommendation information corresponds to a recommendation behavior indicating a history recommendation resource recommended to the object once and having performed an interactive operation by the object, and a plurality of pieces of history recommendation information indicating a history recommendation resource recommended to the object a plurality of times and having performed an interactive operation by the object. For example, the electronic device recommends 100 resources to the object for the first time, and the object performs the interactive operation on only 5 resources of the 100 resources, the 5 resources are used as history recommended resources, the first piece of history recommended information is generated, the first piece of history recommended information indicates the 5 history recommended resources on which the object performs the interactive operation, and after a period of time, the electronic device recommends 50 resources to the object for the second time, and the object performs the interactive operation on only 10 resources of the 50 resources, the 10 resources are used as history recommended resources, and the second piece of history recommended information is generated, the second piece of history recommended information indicates the 10 history recommended resources on which the object performs the interactive operation.

Each piece of historical search information corresponds to one search behavior of the object, and a plurality of pieces of historical recommendation information correspond to a plurality of search behaviors of the object. Each of the history search information indicates a history search resource which is searched based on one history search text and on which the object has performed the interactive operation, and each of the history search information indicates at least one history search resource. For example, the object inputs a first search text through the electronic device, the electronic device searches a plurality of resources based on the first search text, and the object performs an interactive operation on 2 resources among the searched plurality of resources, then the first search text is used as a history search text, the 2 resources are used as history search resources, and a first piece of history search information is generated based on the history search text and the 2 history search resources; and the object inputs a second search text through the electronic equipment, the electronic equipment searches a plurality of resources based on the second search text, and the object performs interactive operation on 3 resources in the searched plurality of resources, so that the second search text is used as a historical search text, the 3 resources are used as historical search resources, and a second piece of historical search information is generated based on the historical search text and the 3 historical search resources.

In some embodiments, each piece of historical recommendation information includes resource information of a historical recommendation resource, a recommendation time of the historical recommendation resource, an interactive operation performed by an object on the historical recommendation resource, and the like. Wherein the recommendation time in the historical recommendation information indicates a time at which the historical recommendation resource was recommended to the subject. The interactive operations in the historical recommendation information indicate which interactive operations the object performs on the historical recommendation resource, and in some embodiments, the interactive operations in the historical recommendation information are represented by operation identifiers of the interactive operations, where the operation identifiers are used to refer to corresponding interactive operations. The resource information of the history recommended resource is used to characterize the history recommended resource, and for example, the resource information includes a resource name, a type to which the resource belongs, profile information of the resource, a publisher, and the like.

In some embodiments, each piece of historical search information includes a search time, historical search text, resource information for a historical search resource, interactive operations performed by an object on the historical search resource, and the like. The search time in the history search information indicates the time of searching the resource based on the history search text, and the history search text is input text when the object searches the resource. The interactive operations in the historical search information indicate which interactive operations the object performed on the historical search resource, and in some embodiments, the interactive operations in the historical search information are represented by operation identifiers of the interactive operations, where the operation identifiers are used to refer to corresponding interactive operations. The resource information of the history search resource is used to characterize the history search resource, and for example, the resource information includes a resource name, a type to which the resource belongs, profile information of the resource, a publisher, and the like.

In step S302, feature extraction is performed on the historical recommendation information and the historical search information, so as to obtain a historical recommendation feature and a historical search feature, where the historical recommendation feature includes a sub-feature for characterizing a historical recommendation resource, and the historical search feature includes a sub-feature for characterizing a historical search text and a historical search resource corresponding to the historical search text.

The historical recommendation features are used for representing historical recommendation information, and the historical search features are used for representing historical search information.

In the embodiment of the disclosure, the historical recommendation feature comprises at least one sub-feature for characterizing the historical recommendation resource, each sub-feature is used for characterizing one historical recommendation resource, and the number of the sub-features in the historical recommendation feature is equal to the number of the historical recommendation resources in the historical recommendation information. Similarly, the history search feature comprises at least one sub-feature for representing the history search text and the history search resource corresponding to the history search text, one sub-feature is used for representing the history search text and the history search resource corresponding to the history search text, and the number of the sub-features in the history search feature is equal to the number of the history search texts in the history search information.

In step S303, the sub-features in the history recommended feature and the sub-features in the history search feature are respectively classified to obtain a first type feature and a second type feature included in the history recommended feature, and a third type feature and a fourth type feature included in the history search feature, where the similarity between the first type feature and the history search feature is not less than a first similarity threshold, the similarity between the second type feature and the history search feature is less than the first similarity threshold, the similarity between the third type feature and the history recommended feature is not less than a second similarity threshold, and the similarity between the fourth type feature and the history recommended feature is less than a second similarity threshold.

In the embodiment of the disclosure, the first type of features and the second type of features are obtained by classifying sub-features in the history recommended features, that is, the first type of features and the second type of features can form the history recommended features, the third type of features and the fourth type of features are obtained by classifying sub-features in the history search features, and the third type of features and the fourth type of features can form the history search features. The sub-features of the history recommended feature and the sub-features of the history search feature are classified so as to determine a first type of feature or a second type of feature which is similar to the history search feature from the history recommended features, and a third type of feature or a fourth type of feature which is similar to the history recommended feature from the history search features, so that the interest of the object can be determined based on the first type of feature, the second type of feature, the third type of feature and the fourth type of feature.

Because the first type of features contained in the history recommended features are similar to the history recommended features, the third type of features contained in the history recommended features are similar to the history recommended features, and then the first type of features and the third type of features are consistent with the interests of the object in the recommended scene and the interests of the object in the search scene, namely, the first type of features and the third type of features can better characterize which resources the object is interested in, the second type of features contained in the history recommended features are dissimilar from the history recommended features, the fourth type of features contained in the history recommended features are dissimilar from the history recommended features, and then the second type of features only correspond to the interests of the object in the recommended scene and the fourth type of features only correspond to the interests of the object in the search scene.

The similarity between the first type of feature and the historical search feature can be determined in any manner, for example, the similarity between the first type of feature and the historical search feature is determined in a euclidean distance manner. The similarity corresponding to the second type of feature, the third type of feature and the fourth type of feature can be determined in any mode.

In some embodiments, the manner in which the first similarity threshold and the second similarity threshold are determined comprises: determining a first sum of the similarities corresponding to the plurality of sub-features in the historical recommended features; determining a ratio between the first sum and the number of the plurality of sub-features in the history recommended feature as a first similarity threshold; determining a second sum of the similarities corresponding to the plurality of sub-features in the historical search feature; the ratio between the second sum and the number of the plurality of sub-features in the historical search feature is determined as a second similarity threshold.

In the embodiment of the disclosure, the history recommendation feature includes a plurality of sub-features, each sub-feature in the history recommendation feature is used for representing one history recommendation resource in the history recommendation information, the history recommendation resources corresponding to different sub-features in the history recommendation feature are different, and the similarity corresponding to any sub-feature in the history recommendation feature is the similarity between the sub-feature and the history search feature. And for the history search feature comprising a plurality of sub-features, each sub-feature in the history search feature is used for representing one history search text in the history search information and the history search resource corresponding to the history search text, the history search texts corresponding to different sub-features in the history search feature are different, and the similarity corresponding to any sub-feature in the history search feature is the similarity of the sub-feature and the history recommendation feature.

In the embodiment of the disclosure, the average value of the similarities corresponding to the plurality of sub-features in the historical recommended features is used as a first similarity threshold value to ensure that the first type of features similar to the historical search features and the second type of features dissimilar to the historical search features can be distinguished from the historical recommended features, so that the accuracy of the determined first type of features and the determined second type of features is ensured.

In the embodiment of the disclosure, the average value of the similarities corresponding to the plurality of sub-features in the historical search features is used as the second similarity threshold value, so that the third type of features similar to the historical recommended features and the fourth type of features dissimilar to the historical search features can be distinguished from the historical search features, and the accuracy of the determined third type of features and fourth type of features is further ensured.

In some embodiments, the similarity corresponding to the plurality of sub-features in the history recommended feature is obtained through normalization, the similarity corresponding to the plurality of sub-features in the history search feature is obtained through normalization, the sum of the similarities corresponding to the plurality of sub-features in the history recommended feature is 1, the sum of the similarities corresponding to the plurality of sub-features in the history search feature is 1, and the determined first similarity threshold is the inverse of the number of the plurality of sub-features in the history recommended feature, and the determined second similarity threshold is the inverse of the number of the plurality of sub-features in the history search feature.

For example, the number of the history recommended resources corresponding to the history recommended features is 20, the sum of the similarities corresponding to the plurality of sub-features in the history recommended features is 1, then 1/20 is determined as a first similarity threshold, the similarity between the first type of features in the history recommended features and the history search features is not less than 1/20, and the similarity between the second type of features in the history recommended features and the history search features is less than 1/20. And determining 1/10 as a second similarity threshold value if the number of the historical search texts corresponding to the historical search features is 10, wherein the similarity between the third type of features in the historical search features and the historical recommended features is not less than 1/10, and the similarity between the fourth type of features in the historical search features and the historical recommended features is less than 1/10.

In some embodiments, this step S303 includes steps 1-3.

Step 1, comparing the historical recommended features with the historical search features to obtain first similarity information and second similarity information, wherein the first similarity information indicates the similarity between each sub-feature in the historical recommended features and the historical search features, and the second similarity information indicates the similarity between each sub-feature in the historical search features and the historical recommended features.

The historical search resources corresponding to the historical search text are resources which are searched based on the historical search text and the objects execute interactive operation. The first similarity information and the second similarity information can each be represented in any form, for example, the first similarity information and the second similarity information can each be represented in the form of feature vectors.

In the embodiment of the disclosure, in the comparison process, each sub-feature in the history recommended feature is compared with the history search feature to obtain the similarity between each sub-feature and the history search feature, and the obtained similarity forms first similarity information; comparing each sub-feature in the historical search feature with the historical recommendation feature to obtain the similarity between each sub-feature and the historical recommendation feature, and forming second similarity information by the obtained similarity.

In some embodiments, the process of obtaining the first similarity information and the second similarity information includes: and fusing the historical recommendation characteristics and the historical search characteristics according to the weights to obtain affinity characteristics, wherein the affinity characteristics indicate the relevance between the historical recommendation characteristics and the historical search characteristics, fusing the historical recommendation characteristics and the affinity characteristics to obtain first similarity information, and fusing the historical search characteristics and the affinity characteristics to obtain second similarity information.

Wherein the weight is constant. In the embodiment of the disclosure, in consideration of the relevance between the historical recommendation feature and the historical search feature, a common Attention mechanism (Co-Attention) is adopted to compare the historical recommendation feature with the historical search feature, so as to obtain first similarity information and second similarity information, and accuracy of the obtained similarity information is guaranteed.

In some embodiments, the first similarity information and the second similarity information satisfy the following relationship:

A＝tanh(H _s W _l (H _r ) ^T )

wherein A is used for representing affinity characteristics, H _s For representing historical search features, H _r For representing historical recommended features, W _l For the weight matrix, tanh (·) is used to represent the hyperbolic tangent function, T is used to represent the transpose of the matrix, a ^r For representing the first similarity information, W _s For representing a constant matrix, softmax (·) for representing a normalization function; a, a ^s For representing the second similarity information, W _r For representing a constant matrix.

And step 2, classifying the sub-features in the history recommended features based on the first similarity information to obtain first-class features and second-class features.

In the embodiment of the disclosure, the first similarity information indicates the similarity between each sub-feature in the history recommended feature and the history search feature, and based on the first similarity information, sub-features with similarity not smaller than a first similarity threshold form a first type feature and sub-features with similarity not smaller than the first similarity threshold form a second type feature from the history recommended feature.

And step 3, classifying the sub-features in the historical search features based on the second similarity information to obtain a third type of features and a fourth type of features.

In the embodiment of the disclosure, the second similarity information indicates a similarity between each sub-feature in the history search feature and the history recommendation feature, and based on the second similarity information, sub-features with a similarity not smaller than a second similarity threshold form a third type feature and sub-features with a similarity smaller than the second similarity threshold form a fourth type feature from the history search features.

In the embodiment of the disclosure, the sub-features in the history search feature and the sub-features in the history recommendation feature are compared so as to respectively classify the sub-features in the history search feature and the sub-features in the history recommendation feature according to the obtained similarity information, so that a first type of feature and a second type of feature which are similar to the history search feature, and a third type of feature and a fourth type of feature which are similar to the history recommendation feature are obtained, and the accuracy of the obtained type of features is ensured.

In step S304, based on the similarity between the resource features of the resource and the historical recommendation features, the first type features and the second type features, the historical recommendation features, the first type features and the second type features are fused to obtain recommendation interest features, and the recommendation interest features indicate the interest degree of the object on the resource in the recommendation scene.

In the embodiment of the disclosure, the resource characteristics of the resource are used for representing the resource, the history recommendation characteristics can reflect the history recommendation resources interested by the object in the recommendation scene, the first type characteristics accord with the interests of the object in the recommendation scene and the search scene, the second type characteristics accord with the interests of the object in the recommendation scene only, the history recommendation characteristics, the first type characteristics and the second type characteristics are fused through the similarity between the resource characteristics of the resource and the history recommendation characteristics, the similarity between the resource to be recommended in the recommendation scene and the history recommendation resources interested by the object in the recommendation scene is considered, the more interested of the object in the history recommendation resources under the influence of the search scene is considered, and the interested degree of the object to be recommended in the recommendation scene can be simulated, so that the accuracy of the determined recommendation interest characteristics is ensured.

In some embodiments, this step S304 includes: and respectively fusing the historical recommendation characteristic, the first class characteristic and the second class characteristic with the resource characteristic of the resource to obtain the similarity between the resource characteristic of the resource and the historical recommendation characteristic, the first class characteristic and the second class characteristic, determining the product of the historical recommendation characteristic and the corresponding similarity, the product of the first class characteristic and the corresponding similarity and the product of the second class characteristic and the corresponding similarity, and splicing the product of the historical recommendation characteristic and the corresponding similarity, the product of the first class characteristic and the corresponding similarity and the product of the second class characteristic and the corresponding similarity to obtain the recommendation interest characteristic.

In the embodiment of the disclosure, the resource features of the resource are respectively fused with the historical recommendation features, the first type features and the second type features, so that the obtained similarity can reflect the relevance of the historical recommendation features, the first type features and the second type features to the resource features of the resource, and further the recommendation interest features can reflect the interest degree of the object on the resource in the recommendation scene, and the accuracy of the recommendation interest features is ensured.

The method for respectively fusing the historical recommendation characteristics, the first type characteristics and the second type characteristics with the resource characteristics of the resources can adopt a multi-head attention mechanism for fusion, namely, adopts a self-attention mechanism to fuse the historical recommendation characteristics with the resource characteristics of the resources to obtain the similarity of the resource characteristics of the resources and the historical recommendation characteristics; adopting a self-attention mechanism to fuse the first type of characteristics with the resource characteristics of the resources to obtain the similarity between the resource characteristics of the resources and the first type of characteristics; and adopting a self-attention mechanism to respectively fuse the second type of characteristics with the resource characteristics of the resource to obtain the similarity between the resource characteristics of the resource and the second type of characteristics.

In step S305, the historical search feature, the third class feature and the fourth class feature are fused based on the similarity between the resource feature of the resource and the historical search feature, the third class feature and the fourth class feature, respectively, to obtain the search interest feature.

In the embodiment of the disclosure, the resource features of the resource are used for representing the resource, the historical search features can reflect the historical search resource which is interested in the object in the search scene, the third type features accord with the interests of the object in the recommendation scene and the search scene, the fourth type features accord with the interests of the object in the search scene only, the resource features, the historical search features, the third type features and the fourth type features of the resource are fused through the similarity between the resource features of the resource and the historical search features, the similarity between the third type features and the fourth type features of the resource respectively, the similarity between the resource to be recommended in the search scene and the historical recommended resource which is interested in the object in the search scene is considered, the historical search resources which are interested in the object in the recommendation scene are considered more, the interest degree of the object to be recommended in the search scene can be simulated, and the accuracy of the determined search interest features is further ensured.

In some embodiments, this step S305 includes: and respectively fusing the historical search feature, the third class feature and the fourth class feature with the resource feature of the resource to obtain the similarity between the resource feature of the resource and the historical search feature, the third class feature and the fourth class feature, determining the product of the historical search feature and the corresponding similarity, the product of the third class feature and the corresponding similarity and the product of the fourth class feature and the corresponding similarity, and splicing the product of the historical search feature and the corresponding similarity, the product of the third class feature and the corresponding similarity and the product of the fourth class feature and the corresponding similarity to obtain the search interest feature.

The method for respectively fusing the historical search features, the third type features and the fourth type features with the resource features of the resources can adopt a multi-head attention mechanism for fusing, namely, adopts a self-attention mechanism to fuse the historical search features with the resource features of the resources to obtain the similarity of the resource features of the resources and the historical search features; adopting a self-attention mechanism to fuse the third type of characteristics with the resource characteristics of the resources to obtain the similarity between the resource characteristics of the resources and the third type of characteristics; and adopting a self-attention mechanism to respectively fuse the fourth type of characteristics with the resource characteristics of the resource to obtain the similarity between the resource characteristics of the resource and the fourth type of characteristics.

It should be noted that, in the embodiment of the present disclosure, taking the history recommendation feature corresponding to the history recommendation information and the history search feature corresponding to the history search information as an example, the history recommendation feature and the history search feature are used to obtain the recommendation interest feature and the search history feature, and in another embodiment, the steps S302-S305 are not required to be executed, but other manners are adopted to obtain the recommendation interest feature and the search interest feature of the object based on the resource to be recommended, the history recommendation information and the history search information.

In step S306, the object features of the object, the resource features of the resource, the recommended interest features, and the search interest features are spliced to obtain spliced features.

In the embodiment of the disclosure, the object feature is used for representing the object, the resource feature is used for representing the resource, the recommended interest feature and the search feature can reflect the interest degree of the object on the resource under the recommended or searched scene respectively, and the object feature, the resource feature, the recommended interest feature and the search interest feature are spliced to enrich the information content contained in the spliced feature, so that the recommendation probability can be predicted based on the spliced feature.

When the object feature, the resource feature, the recommended interest feature and the search interest feature are spliced, the object feature, the resource feature, the recommended interest feature and the search interest feature can be spliced in any order, which is not limited in the disclosure.

In some embodiments, the object features are feature extraction of object information of the object. Wherein the object information is used to characterize the object. For example, the object is a user, and the object information is user information.

In step S307, the stitching features are subjected to dimension reduction processing, so as to obtain a recommendation probability of the resource, where the recommendation probability indicates a possibility that the object performs an interactive operation on the resource when recommending the resource to the object.

For example, the stitching feature is a multidimensional vector, such as a vector with stitching feature of 1×64, and the stitching feature is transformed to reduce the number of dimensions contained in the stitching feature, so as to obtain a value, i.e. the recommendation probability.

It should be noted that, in the embodiment of the present disclosure, the recommendation probability of any resource to be recommended is merely taken as an example to describe, and according to the above steps S304 to S307, the recommendation probability of a plurality of resources to be recommended can be obtained.

It should be noted that, in the embodiment of the present disclosure, the object feature of the object, the resource feature of the resource, the recommended interest feature and the search interest feature are spliced as an example, and the recommended probability is obtained by using the spliced feature, and in another embodiment, the steps S306-S307 are not required to be executed, but other modes are adopted to process the object feature of the object, the resource feature of the resource, the recommended interest feature and the search interest feature, so as to obtain the recommended probability of the resource.

In step S308, a target resource among the plurality of resources is recommended to the object based on the recommendation probability of the plurality of resources to be recommended.

In some embodiments, this step S308 includes the following two approaches.

According to the first mode, based on the recommendation probability of the plurality of resources to be recommended, a first number of target resources with the maximum recommendation probability are selected from the plurality of resources to be recommended, and the selected target resources are recommended to the object.

Wherein the first number is any number, e.g., the first number is 3 or 5, etc.

In the embodiment of the disclosure, for a plurality of resources to be recommended, the recommendation probability of the selected target resource is larger than the recommendation probability of the unselected resource, so as to ensure that the recommendation probability of the selected target resource is large enough, further ensure that the selected target resource is the resource of interest to the object, and further ensure the accuracy of resource recommendation.

Taking the first number as 3 as an example, the number of the plurality of resources to be recommended at this time is 20, and under the condition that the recommendation probability of each resource to be recommended is obtained, sequencing the plurality of resources with the recommendation probability of 20 resources from big to small, and recommending the 3 resources with the front sequencing to the object.

And selecting a target resource with the recommendation probability larger than a probability threshold from the plurality of resources to be recommended based on the recommendation probability of the plurality of resources to be recommended, and recommending the selected target resource to the object.

The probability threshold is an arbitrary value, and the recommendation probability of any resource is larger than the probability threshold, so that the object is interested in the resource sufficiently, and therefore, only the resource with the recommendation probability larger than the probability threshold is recommended to the object, the resources recommended to the object are all interested in the object, and the accuracy of the resource recommendation is further ensured.

It should be noted that, in the embodiment of the present disclosure, the case where the recommendation probability is greater than the probability threshold is taken as an example to describe the embodiment, and in another embodiment, in the case where the recommendation probability is not greater than the probability threshold, it means that, in the case where the resource is recommended to the object, the possibility that the object performs the interactive operation on the resource is small, and it is reflected that the object is not interested in the resource, and the resource is no longer recommended to the object.

It should be noted that, in this disclosure, only any resource to be recommended is taken as an example for illustration, and in another embodiment, a plurality of resources to be recommended are obtained, after step S303, a recommendation probability corresponding to each resource in the plurality of resources to be recommended is obtained according to steps S304 to S307, a resource with a recommendation probability greater than a probability threshold is selected from the plurality of resources to be recommended, and the selected resource is recommended to the object, so that the recommendation probability corresponding to each resource recommended to the object is greater than the probability threshold.

In the embodiment of the disclosure, the resource characteristics of the resource are used for representing the resource, the history recommendation characteristics can reflect the history recommendation resource which is interested by the object in the recommendation scene, the first type characteristics accord with the interests of the object in the recommendation scene and the search scene, the second type characteristics accord with the interests of the object in the recommendation scene, and the similarity between the resource characteristics, the history recommendation characteristics, the first type characteristics and the second type characteristics of the resource to be recommended in the recommendation scene and the history recommendation resource which is interested by the object in the recommendation scene is considered, and the influence of the search scene on which history recommendation resource is more interested by the object is considered, so that the interest degree of the object to be recommended in the recommendation scene can be simulated, and the accuracy of the determined recommendation interest characteristics is further ensured.

In the embodiment of the disclosure, the resource features of the resource are used for representing the resource, the historical search features can reflect the historical search resource which is interested in the object in the search scene, the third type features accord with the interests of the object in the recommendation scene and the search scene, and the fourth type features accord with the interests of the object in the search scene only, and by fusing the resource features, the historical search features, the third type features and the fourth type features of the resource, the similarity between the resource to be recommended in the search scene and the historical recommended resource which is interested in the object in the search scene is considered, the historical search resources which are interested in the object in the recommendation scene are considered, the interest degree of the object to be recommended in the search scene is simulated, and the accuracy of the determined search interest features is further ensured.

On the basis of the embodiment shown in fig. 3, the embodiment of the disclosure may also take the example that the history recommendation information includes a plurality of history recommendation resources and the history search information includes a plurality of history search texts, obtain the history search feature based on the plurality of history search resources, and obtain the history recommendation feature based on the plurality of history search texts and the history search resource corresponding to each history search text, which is described in detail in the following embodiments.

FIG. 4 is a flowchart illustrating a resource recommendation method, as shown in FIG. 4, performed by an electronic device, according to an exemplary embodiment, comprising the steps of:

in step S401, feature extraction is performed on each history recommended resource in the history recommended information, so as to obtain a resource feature of each history recommended resource, and the resource feature of each history recommended resource forms a history recommended feature.

The resource characteristics of each history recommended resource are used for representing the corresponding history recommended resource, the resource characteristics of each history recommended resource can be represented in any form, for example, the resource characteristics of each history recommended resource are represented in the form of feature vectors, the history recommended characteristics formed by the resource characteristics of a plurality of history recommended resources included in the history recommended information are represented in the form of feature matrices, and each row of feature vectors in the feature matrices is the resource characteristics of one history recommended resource. And (3) mapping the history recommended resources in the highly discrete space into a dense feature space by adopting a feature extraction mode to obtain resource features of the history recommended resources.

In some embodiments, each piece of history recommendation information indicates one or more history recommendation resources, and then the resource features of the history recommendation resources indicated by the pieces of history recommendation information are formed into history recommendation features.

In some embodiments, before the step S401, repeated historical recommendation resources in the historical recommendation information are filtered, that is, the method further includes: and filtering repeated historical recommendation resources in the historical recommendation information.

In the embodiment of the disclosure, the object may have one or more pieces of history recommendation information, each piece of history recommendation information indicates one or more history recommendation resources, and since different history recommendation information may have the same history recommendation resources or the same history recommendation resources in the same history recommendation information, repeated history recommendation resources in the history recommendation information are filtered, so that the filtered history recommendation resources do not have repeated resources, so that the history recommendation characteristics are formed based on the resource characteristics of the filtered history recommendation resources, and the accuracy of the history recommendation characteristics is further ensured.

In some embodiments, the process of obtaining the resource characteristics of each of the historical recommended resources includes: and for any historical recommended resource, extracting the characteristic of the resource information of the historical recommended resource to obtain the resource characteristic of the historical recommended resource.

The resource information includes a resource name, a type to which the resource belongs, profile information of the resource, and the like. In some embodiments, the resource information of the historical recommended resources further includes a number, the code being used to distinguish between different historical recommended resources in the historical recommended information. Wherein the code can be represented in any form, e.g. the number is represented in numerical form. For example, the codes of the plurality of historical recommended resources are 1, 2, 3, etc. In the embodiment of the disclosure, a code is set for each historical recommended resource in the historical recommended information, so that different historical recommended resources in the historical recommended information can be distinguished, extracted resource features can be distinguished, and the accuracy of the resource features is further ensured.

In some embodiments, the process of constructing the resource characteristics of each history recommended resource into history recommended characteristics includes the following steps 1-3:

step 1, acquiring a first position feature of each historical recommended resource, wherein the first position feature indicates a relative time sequence between the historical recommended resource and other historical recommended resources in the historical recommended information.

Wherein the first position feature can be represented in any form, for example, the first position feature is represented in the form of a feature vector. In the plurality of historical recommended resources included in the historical recommended information, each historical recommended resource is a resource which has been recommended to the object before, and the time when different historical recommended resources are recommended to the object may be different, and a first position feature is set for each historical recommended resource, so that the time sequence of recommending the plurality of historical recommended resources to the object can be determined through the first position features of the plurality of historical recommended resources.

In some embodiments, step 1 comprises: according to the recommendation time sequence corresponding to the plurality of historical recommendation resources in the historical recommendation information, setting a sequence number for each historical recommendation resource, and extracting features from the sequence numbers corresponding to each historical recommendation resource to obtain a first position feature of each historical recommendation resource.

The serial number can be represented by an arbitrary character string. For example, according to the sequence from late to early of the recommended time corresponding to the plurality of historical recommended resources, a sequence number 0001 is set for the first historical recommended resource, a sequence number 0002 is set for the second historical recommended resource, and so on, a sequence number is set for each historical recommended resource, wherein the later the recommended time is, the closer the recommended time is to the current time.

And 2, fusing the resource characteristics of each historical recommended resource with the first position characteristics to obtain fusion characteristics corresponding to each historical recommended resource.

In the embodiment of the disclosure, the first position features of the plurality of historical recommended resources can reflect the recommendation time sequence of the plurality of historical recommended resources, and for each historical recommended resource, the resource features of the historical recommended resources are fused with the corresponding first position features to enrich the information contained in the obtained fusion features, so that the fusion features of the plurality of historical recommended resources can reflect the recommendation time sequence of the plurality of historical recommended resources, and also can reflect the time-varying condition of the historical recommended resources of interest of the object, thereby ensuring the accuracy of the fusion features.

And 3, updating the fusion characteristics corresponding to each historical recommended resource based on the fusion characteristics corresponding to the historical recommended resources, and forming the historical recommended characteristics by updating the characteristics of the historical recommended resources.

In some embodiments, this step 3 comprises: and determining the product of the fusion characteristics corresponding to each historical recommended resource and the fusion characteristics corresponding to the first historical recommended resource as the weight corresponding to each historical recommended resource, fusing the fusion characteristics corresponding to the plurality of historical recommended resources based on the weight corresponding to each historical recommended resource, and fusing the fused characteristics with the fusion characteristics corresponding to the first historical recommended resource to obtain the updated characteristics of the first historical recommended resource.

The first historical recommended resource is any one of a plurality of historical recommended resources. The above description only uses the first historical recommended resource as an example, and according to the above manner, the fusion feature corresponding to each historical recommended resource is updated to obtain the updated feature of each historical recommended resource.

In the embodiment of the disclosure, a self-attention mechanism is adopted to update the fusion features corresponding to the plurality of historical recommended resources so as to enhance the relevance among the fusion features corresponding to the plurality of historical recommended resources and further ensure the accuracy of the historical recommended features.

In step S402, feature extraction is performed on the history search text and the history search resource in the history search information, so as to obtain text features of each history search text and resource features of each history search resource.

The text features of the historical search text are used for representing the historical search text, the resource features of the historical search resources are used for representing the historical search resources, and the text features and the resource features can be represented in any form, for example, the text features and the resource features can be represented in the form of feature vectors.

In some embodiments, the history search information may include one or more history search texts and one or more history search resources corresponding to each history search text, and feature extraction is performed on each history search text and each history search resource in the history search information to obtain text features of each history search text and resource features of each history search resource.

In some embodiments, before the step S401, repeated historical search text in the historical search information is filtered, that is, the method further includes: filtering repeated historical search texts in the historical search information, and filtering repeated historical search resources in the historical search resources corresponding to the same historical search text.

In the embodiment of the disclosure, the object may have one or more pieces of history search information, each piece of history search information indicates one or more pieces of history search text, and since different pieces of history search information may have the same history search text, or the same history search text exists in the same history search information, repeated history search text in the history search information is filtered to avoid the situation that the history search text is repeated, and the filtered history search text is filtered for repeated history search resources in history search resources corresponding to the same history search text to avoid the situation that the history search resources are repeated, so that the subsequent text features based on the filtered history search text and the resource features of the corresponding history search resources form the history search features, thereby ensuring the accuracy of the history search features.

In some embodiments, the process of obtaining text features for each historical search text includes: and for any historical search text, extracting the characteristics of the historical search text to obtain the text characteristics of the historical search text.

In some embodiments, the process of obtaining the resource characteristics of each historical search resource includes: and for any historical search resource, extracting the characteristic of the resource information of the historical search resource to obtain the resource characteristic of the historical search resource.

The resource information includes a resource name, a type to which the resource belongs, profile information of the resource, and the like. In some embodiments, the resource information of the historical search resources further includes a number, the number being encoded to distinguish between different historical search resources in the historical search information. Wherein the code can be represented in any form, e.g. the number is represented in numerical form. For example, the codes of the plurality of historical search resources are 1, 2, 3, etc. In the embodiment of the disclosure, a code is set for each historical search resource in the historical search information, so that different historical search resources in the historical search information can be distinguished, the extracted resource features can also distinguish different historical search resources, and the accuracy of the resource features is further ensured.

In step S403, the text feature of each history search text and the resource feature of the corresponding history search resource are fused to obtain a first fusion feature corresponding to each history search text, and the first fusion feature corresponding to each history search text forms a history search feature.

In the embodiment of the disclosure, for each history search text and the history search resource corresponding to the history search text, the interest of the object in the history search can be reflected, so that for any history search text, the text feature of the history search text and the feature of the history search resource corresponding to the history search text are fused to obtain a first fusion feature corresponding to the history search text, the first fusion feature is used for representing the interest expression of the object in one history search, and further the first fusion features corresponding to a plurality of history search texts form the history search feature, so that the history search feature can represent the interest of the object in a search scene.

One or more historical search resources are corresponding to any historical search text, and the historical search resources corresponding to the historical search text are resources which are searched based on the historical search text and are subjected to interactive operation by the object. The text feature or the resource feature can be represented in any form, for example, the text feature or the resource feature is represented in the form of a feature vector, then the first fusion feature is also represented in the form of a feature vector, the history search feature is represented in the form of a feature matrix, and each row of feature vector in the feature matrix is the first fusion feature corresponding to one history search text.

In some embodiments, the first fused feature corresponding to each historical search text is formed into a historical search feature, including the following steps 1-3.

Step 1, acquiring a second position feature and a search type feature of each historical search text, wherein the second position feature indicates a relative time sequence between the historical search text and other historical search texts in the historical search information, and the search type feature indicates a search type adopted when searching based on the historical search text.

Wherein the second position feature can be represented in any form, e.g. the second position feature is represented in the form of a feature vector. When each history search text included in the history search information is a text input by the object, and the input time of different history search texts may be different, that is, the search time corresponding to each history search text is different, and a second position feature is set for each history search text, the time sequence of searching resources by the object based on the plurality of history search texts can be determined through the second position features of the plurality of history search texts.

In the embodiment of the disclosure, the search type features indicate the search types adopted when the object searches based on the historical search text, and different search types indicate different search entries, so that the search based on the historical search text is reflected by which search entry the object searches from. When searching is performed from different search portals, different intentions of the object can be reflected, and further different interests of the object can be reflected. For example, when searching from a search portal in an application home page, that is, when an object wants to search based on search text, and when searching from a search portal in a comment interface, that is, when an object wants to search for resources related to comment information in the comment interface based on search text.

In some embodiments, the process of obtaining the second location feature for each historical search text includes: according to the sequence of the search time corresponding to the plurality of historical search texts in the historical search information, setting a sequence number for each historical search text, and extracting features from the sequence numbers corresponding to each historical search text to obtain the second position feature of each historical search text.

The serial number can be represented by an arbitrary character string. For example, in the order of the search times corresponding to the plurality of history search texts from late to early, a sequence number 0001 is set for the first history search text, a sequence number 0002 is set for the second history search text, and so on, a sequence number is set for each history search text, wherein the later the search time is, the closer the search time is to the current time.

In some embodiments, the process of obtaining search type features for each historical search text includes: and acquiring a search type identifier corresponding to any historical search text from the historical search information, and determining the search type characteristic corresponding to the search type identifier as the search type characteristic of the historical search text.

The search type identifier can be represented in any form, for example, in the form of a character string. For the mode of acquiring the search type feature corresponding to the search type identifier, the search type identifier can be obtained by feature extraction, or the search type feature corresponding to each search type identifier is configured, and the search type feature corresponding to the search type identifier is acquired from the configured search type features.

And 2, fusing the first fusion feature, the second position feature and the search type feature corresponding to each historical search text to obtain a second fusion feature corresponding to each historical search text.

In the embodiment of the disclosure, the second position features of the plurality of historical search texts can reflect the search time sequence of the plurality of historical search texts, the search type features indicate the search type adopted when searching is performed based on the historical search texts, and for each historical search text, the first fusion features corresponding to the historical search texts are fused with the corresponding first position features and search type features so as to enrich the information contained in the obtained second fusion features, so that the second fusion features of the plurality of historical search texts can reflect the search time sequence of the plurality of historical search texts and the search type adopted, and the interest of the object can also be reflected in time change, thereby ensuring the accuracy of the second fusion features.

And 3, updating the second fusion features corresponding to each historical search text based on the second fusion features corresponding to the plurality of historical search texts, and forming the historical search features by the updated features of the plurality of historical search texts.

In some embodiments, this step 3 comprises: and determining the product of the second fusion feature corresponding to each historical search text and the second fusion feature corresponding to the first historical search text as the weight corresponding to each historical search text, fusing the second fusion features corresponding to the plurality of historical search texts based on the weight corresponding to each historical search text, and fusing the fused features with the second fusion features corresponding to the first historical search text to obtain updated features of the first historical search text.

The first historical search text is any one of a plurality of historical search texts. The above description is given by taking the first historical search text as an example, and according to the above manner, the fusion feature corresponding to each historical search text is updated to obtain the updated feature of each historical search text.

In the embodiment of the disclosure, a self-attention mechanism is adopted to update the fusion features corresponding to the plurality of historical search texts so as to enhance the relevance among the fusion features corresponding to the plurality of historical search texts and further ensure the accuracy of the historical recommendation features.

It should be noted that, on the basis of the embodiments shown in fig. 2 to 4, the embodiments of the present disclosure can also call the recommendation prediction model to perform the process of obtaining the recommendation probability, that is, call the recommendation prediction model to perform the steps S202 to S203, or perform the steps S302 to S307, or perform the steps S401 to S403. Before invoking the recommendation prediction model to obtain the recommendation probability of any resource for any object, the recommendation prediction model needs to be trained, and the training process is described in the following embodiments.

FIG. 5 is a flowchart illustrating a recommended prediction model training method, as shown in FIG. 5, performed by an electronic device, according to an exemplary embodiment, comprising the steps of:

in step S501, a sample object, a sample resource, a sample recommendation probability of the sample resource, sample history recommendation information of the sample object, and sample history search information are acquired, the sample history recommendation information indicating a history recommendation resource in which the sample object has performed an interactive operation, the sample history search information indicating a history search text input by the sample object and a history search resource which has been searched based on the history search text and in which the sample object has performed an interactive operation.

The sample recommendation probability is whether the sample object performs an interactive operation on the sample resource under the condition that the sample resource is recommended to the sample object. In some embodiments, the sample recommendation probability is 0 or 1;0 indicates that in the case of recommending a sample resource to a sample object, the sample object does not perform an interactive operation on the sample resource; 1 denotes that in case of recommending a sample resource to a sample object, the sample object performs an interactive operation on the sample resource.

In some embodiments, the sample resources and sample recommendations for the sample resources are determined based on historical recommendations or historical search records for the sample objects.

The history recommendation record indicates history recommended resources recommended to the sample object, and indicates which history recommended resources the sample object performs interactive operation on, and which history recommended resources do not perform interactive operation on. The history search record indicates the history search text input by the sample object, the history search resources searched based on the history search text, and indicates which history search resources the sample object performs interactive operation on, and which history search resources do not perform interactive operation on. Sample history recommendation information for the sample object is determined based on the sample object's history recommendation record, and sample history search information for the sample object is determined based on the sample object's history search record.

The sample resource is any resource in the history recommending record or the history searching record, and the sample recommending probability of the sample resource is determined based on the condition that the sample object in the history recommending record or the history searching record executes the interactive operation on the sample resource. In the case of determining the sample resource, the history recommendation information is determined from the history recommendation record based on the history search time or the history recommendation time corresponding to the sample resource, and the history search information is determined from the history search record.

For example, taking a sample resource as any resource in a history recommendation record, determining a history recommendation time corresponding to the sample resource, and generating history recommendation information based on the history recommendation resource of which the recommendation time is before the history recommendation time in the history recommendation record and the sample object has executed an interactive operation; and determining a historical search text with the search time before the historical recommendation time from the historical search record, and generating historical search information based on the historical search resources corresponding to the determined historical search text, the historical search resources with the sample objects subjected to the interaction operation and the determined historical search text.

In step S502, a recommendation prediction model to be trained is invoked, and based on sample resources, sample history recommendation information and sample history search information, a recommendation interest feature and a search interest feature of a sample object are obtained, wherein the recommendation interest feature indicates an interest degree of the sample object to the sample resources in a recommendation scene, and the search interest feature indicates an interest degree of the sample object to the sample resources in a search scene.

The recommended prediction model to be trained is an arbitrary network model, for example, the recommended prediction model to be trained is SESRec (a Search-Enhanced framework for Sequential Recommendation, a model of an auxiliary lifting sequence recommendation model enhanced based on Search data).

The step S502 is similar to the step S202, and will not be described again.

In step S503, a recommendation prediction model to be trained is invoked, and dimension reduction processing is performed on object features of the sample object, resource features of the sample resource, recommendation interest features and search interest features, so as to obtain a prediction recommendation probability of the sample resource, where the prediction recommendation probability indicates a possibility that the sample object performs an interactive operation on the sample resource when recommending the sample resource to the sample object.

The step S503 is similar to the step S203, and will not be described again.

In step S504, training the recommended prediction model to be trained based on the sample recommendation probability and the predicted recommendation probability, to obtain a target recommended prediction model.

In the embodiment of the disclosure, the sample recommendation probability can reflect the real situation that the sample object performs the interactive operation on the sample resource, and the prediction recommendation probability is based on the situation that the sample object predicted by the recommendation prediction model performs the interactive operation on the sample resource, and the difference between the sample recommendation probability and the prediction recommendation probability can reflect the accuracy of the recommendation prediction model, and the recommendation prediction model is trained based on the sample recommendation probability and the prediction recommendation probability so as to improve the accuracy of the recommendation prediction model to be trained.

The target recommendation prediction model is a trained recommendation prediction model, and the target recommendation prediction model is called to execute the process of acquiring recommendation probability according to the embodiment shown in fig. 2 to 4.

According to the scheme provided by the embodiment of the disclosure, the historical recommendation information and the historical search information are considered to respectively represent which resources are interested by the object in the recommendation scene and the search scene, the historical search information is taken as auxiliary information in the recommendation scene, the historical recommendation information and the historical search information are combined, the degree of interest of the object to be recommended resources in the recommendation scene and the search scene can be simulated, the sample recommendation probability of the sample object, the sample resource, the sample historical recommendation information of the sample resource and the sample historical search information are combined, and the recommendation prediction model to be trained is trained, so that the accuracy of the recommendation prediction model is improved, the recommendation probability of the object to perform interactive operation on the resources under the condition that any resource is recommended to any object can be obtained by utilizing the target recommendation prediction model obtained through training, and further the resource recommendation effect is ensured.

In some embodiments, the recommended interest feature of the sample object is obtained by fusing a resource feature, a historical recommendation feature of sample historical recommendation information, a first type feature and a second type feature contained in the historical recommendation feature, wherein the similarity between the first type feature and the historical search feature of the sample historical search information is not less than a first similarity threshold, the similarity between the second type feature and the historical search feature is less than the first similarity threshold, the first type feature and the second type feature are obtained by classifying sub-features in the historical recommendation feature, and the sub-features in the historical recommendation feature are used for representing historical recommendation resources; based on the sample recommendation probability and the prediction recommendation probability, training the recommendation prediction model to be trained, and before obtaining the target recommendation prediction model, the method further comprises the following steps:

Determining a first similarity of the historical recommended features and the first type of features and a second similarity of the historical recommended features and the second type of features;

training a recommendation prediction model to be trained based on the sample recommendation probability and the prediction recommendation probability to obtain a target recommendation prediction model, comprising:

based on the sample recommendation probability, the prediction recommendation probability, the first similarity and the second similarity, training the recommendation prediction model to be trained so that the first similarity is increased and the second similarity is reduced, and obtaining the target recommendation prediction model.

In the embodiment of the disclosure, the historical recommendation feature, the first type feature and the second type feature are all obtained based on a recommendation prediction model, the first type feature and the second type feature are obtained by classifying sub-features in the historical recommendation feature, the similarity between the first type feature and the historical search feature is not smaller than a first similarity threshold, the similarity between the second type feature and the historical search feature is smaller than the first similarity threshold, and the recommendation prediction model is trained based on the first similarity and the second similarity, so that the first similarity is increased and the second similarity is reduced, that is, the historical recommendation feature obtained based on the recommendation prediction model is more similar to the first type feature, and the historical recommendation feature is more dissimilar to the second type feature, so that the historical recommendation feature obtained based on the recommendation prediction model can reflect the interest of a sample object more, and the accuracy of the recommendation prediction model is improved.

In some embodiments, the search interest feature of the sample object is obtained by fusing a resource feature, a historical search feature of sample historical search information, a third type feature and a fourth type feature contained in the historical search feature, the similarity between the third type feature and the historical recommendation feature of the sample historical recommendation information is not less than a second similarity threshold, the similarity between the fourth type feature and the historical recommendation feature is less than the second similarity threshold, the third type feature and the fourth type feature are obtained by classifying sub-features in the historical search feature, and the sub-features in the historical search feature are used for representing historical search texts and historical search resources corresponding to the historical search texts; based on the sample recommendation probability and the prediction recommendation probability, training the recommendation prediction model to be trained, and before obtaining the target recommendation prediction model, the method further comprises the following steps:

determining a third similarity of the historical search feature and the third type of feature and a fourth similarity of the historical search feature and the fourth type of feature;

based on the sample recommendation probability, the prediction recommendation probability, the third similarity and the fourth similarity, training the recommendation prediction model to be trained so that the third similarity is increased and the fourth similarity is reduced, and obtaining the target recommendation prediction model.

In the embodiment of the disclosure, the historical search feature, the third type feature and the fourth type feature are all obtained based on a recommendation prediction model, the third type feature and the fourth type feature are obtained by classifying sub-features in the historical recommendation feature, the similarity between the third type feature and the historical recommendation feature is not smaller than a second similarity threshold, the similarity between the fourth type feature and the historical recommendation feature is smaller than the second similarity threshold, and the recommendation prediction model is trained based on the third similarity and the fourth similarity so that the third similarity is increased and the fourth similarity is reduced, that is, the historical search feature obtained based on the recommendation prediction model is more similar to the third type feature, and the historical search feature is not similar to the fourth type feature, so that the historical search feature obtained based on the recommendation prediction model can reflect the interest of a sample object more, and the accuracy of the recommendation prediction model is improved.

In some embodiments, the history search features are formed by fusion features corresponding to each history search text in the sample history search information, and the fusion features corresponding to the history search text are obtained by fusion of text features of the history search text and features of corresponding history search resources; based on the sample recommendation probability, the prediction recommendation probability, the third similarity and the fourth similarity, before training the recommendation prediction model to be trained, the method further comprises:

Determining a fifth similarity of text features of the historical search text and features of the corresponding historical search resources;

training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity and the fourth similarity so that the third similarity is increased and the fourth similarity is decreased, and obtaining the target recommendation prediction model comprises the following steps:

and training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity and the fifth similarity so that the third similarity is increased, the fourth similarity is reduced and the fifth similarity is increased, and obtaining the target recommendation prediction model.

In the embodiment of the disclosure, the history search resource corresponding to the history search text is searched based on the history search text, and then the similarity of the history search text and the corresponding history search resource is reflected. The fifth similarity is the text feature of the historical search text and the feature of the corresponding historical search resource, and the text feature and the feature of the historical search resource are obtained based on the recommendation prediction model, so that the fifth similarity can reflect the accuracy of the text feature obtained based on the recommendation prediction model and the feature of the historical search resource, further reflect the accuracy of the recommendation prediction model, and the recommendation prediction model is trained based on the fifth similarity, so that the fifth similarity is increased, and the accuracy of the recommendation prediction model is improved.

In some embodiments, prior to training the recommended prediction model to be trained, the method further comprises:

determining a negative sample resource of a historical search text and a negative sample text of the historical search resource corresponding to the historical search text from the sample historical search information, wherein the negative sample resource is any historical search resource except the historical search resource corresponding to the historical search text in the sample historical search information, and the negative sample text is any historical search text except the historical search text in the sample historical search information;

determining a sixth similarity of the historical search text and the negative sample text, and a seventh similarity of the historical search resource corresponding to the historical search text and the negative sample text;

training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity, the sixth similarity and the seventh similarity so as to increase the third similarity, decrease the fourth similarity, decrease the sixth similarity and decrease the seventh similarity, and obtaining the target recommendation prediction model.

In the embodiment of the disclosure, the history search text and the corresponding history search resource are respectively used as the corresponding positive samples, the negative samples are determined for the history search text and the corresponding history search resource, the similarity corresponding to the positive samples and the similarity corresponding to the negative samples are utilized to train the recommendation prediction model, so that the distances between the history search text and the positive sample resource and the distances between the history search resource and the positive sample text in the feature space are shortened, the distances between the history search text and the negative sample resource and the distances between the history search resource and the negative sample text in the feature space are shortened, a self-supervision training mode is realized, and the accuracy of the recommendation prediction model is improved.

Based on the embodiment shown in fig. 5, the disclosed embodiment can also take various losses to train the recommended prediction model, the training process is detailed in the following embodiments.

FIG. 6 is a flowchart illustrating a recommended prediction model training method, as shown in FIG. 6, performed by an electronic device, according to an exemplary embodiment, comprising the steps of:

in step S601, a sample object, a sample resource, a sample recommendation probability of the sample resource, sample history recommendation information of the sample object, and sample history search information are acquired, the sample history recommendation information indicates a history recommendation resource in which the sample object has performed an interactive operation, and the sample history search information indicates a history search text input by the sample object and a history search resource which has been searched based on the history search text and in which the sample object has performed an interactive operation.

The step S601 is similar to the step S501, and will not be described again.

In step S602, a recommendation prediction model to be trained is called, and feature extraction is performed on the sample historical recommendation information and the sample historical search information, so as to obtain historical recommendation features of the sample historical recommendation information and historical search features of the sample historical search information.

The step S602 is similar to the step S302, and will not be described herein.

In step S603, a recommendation prediction model to be trained is invoked, sub-features in the history recommendation features and sub-features in the history search features are classified, and a first type of features and a second type of features contained in the history recommendation features and a third type of features and a fourth type of features contained in the history search features are obtained, the similarity between the first type of features and the history search features is not smaller than a first similarity threshold, the similarity between the second type of features and the history search features is smaller than the first similarity threshold, the similarity between the third type of features and the history recommendation features is not smaller than a second similarity threshold, and the similarity between the fourth type of features and the history recommendation features is smaller than a second similarity threshold.

The step S603 is similar to the step S303, and will not be described again.

In step S604, a recommendation prediction model to be trained is invoked, and the resource features, the history recommendation features, the first type features and the second type features of the sample resource are fused based on the similarity between the resource features of the sample resource and the history recommendation features, the first type features and the second type features, respectively, to obtain recommended interest features of the sample object, wherein the recommended interest features indicate the interest degree of the sample object on the sample resource in a recommendation scene.

The step S604 is similar to the step S304, and will not be described again.

In step S605, the resource feature, the history search feature, the third class feature and the fourth class feature of the sample resource are fused based on the similarity between the resource feature and the history search feature, the third class feature and the fourth class feature, respectively, of the sample resource, so as to obtain a search interest feature of the sample object, wherein the search interest feature indicates the interest degree of the sample object on the sample resource in the search scene.

The step S605 is similar to the step S305, and will not be described again.

It should be noted that, in the embodiment of the present disclosure, taking the history recommendation feature corresponding to the obtained history recommendation information and the history search feature corresponding to the history search information as an example, the history recommendation feature and the history search feature are used to obtain the recommendation interest feature and the search history feature, and in another embodiment, the recommendation prediction model is called in other manners without executing the steps S602 to S605, and the recommendation interest feature and the search interest feature of the sample object are obtained based on the sample resource, the sample history recommendation information and the sample history search information.

In step S606, a recommended prediction model to be trained is called, and object features of the sample object, resource features of the sample resource, recommended interest features and search interest features are spliced to obtain sample splicing features.

In step S607, a recommendation prediction model to be trained is invoked, and dimension reduction processing is performed on the sample stitching feature, so as to obtain a prediction recommendation probability of the sample resource, where the prediction recommendation probability indicates a possibility that the sample object performs an interactive operation on the sample resource when the sample resource is recommended to the sample object.

The steps S606-S607 are similar to the steps S306-S307, and are not described herein.

It should be noted that, in the embodiment of the present disclosure, the object feature of the sample object, the resource feature of the sample resource, the recommended interest feature and the search interest feature are taken as examples, and the predicted recommended probability is obtained by using the sample stitching feature, and in another embodiment, the steps S606-S607 are not required to be executed, but other modes are adopted to process the object feature of the sample object, the resource feature of the sample resource, the recommended interest feature and the search interest feature, so as to obtain the predicted recommended probability of the sample resource.

In step S608, training the recommendation prediction model to be trained based on the sample recommendation probability and the prediction recommendation probability, to obtain a target recommendation prediction model.

In some embodiments, this step S608 includes: and determining a first loss value based on the sample recommendation probability and the prediction recommendation probability, and training a recommendation prediction model based on the first loss value.

In the embodiment of the disclosure, a cross entropy loss mode is adopted, a first loss value is determined, and a recommended prediction model is trained based on the first loss value, so that accuracy of the recommended prediction model is improved.

In some embodiments, this step S607 includes the following three ways.

The first way is: determining a first similarity of the historical recommended features and the first type of features and a second similarity of the historical recommended features and the second type of features; based on the sample recommendation probability, the prediction recommendation probability, the first similarity and the second similarity, training the recommendation prediction model to be trained so that the first similarity is increased and the second similarity is reduced, and obtaining the target recommendation prediction model.

In the embodiment of the disclosure, the recommended interest feature of the sample object is obtained by fusing a resource feature of the sample resource, a historical recommendation feature of sample historical recommendation information, a first type feature and a second type feature contained in the historical recommendation feature, wherein the similarity between the first type feature and the historical search feature of the sample historical search information is not less than a first similarity threshold, and the similarity between the second type feature and the historical search feature is less than the first similarity threshold.

The first similarity is used for representing the similarity degree of the historical recommended features and the first type of features, and the second similarity is used for representing the similarity degree of the historical recommended features and the second type of features. The first similarity and the second similarity can be obtained in any manner, for example, in such a manner that the euclidean distance is used to obtain the first similarity and the second similarity.

In some embodiments, the first way comprises: and determining a first loss value based on the sample probability and the prediction recommendation probability, determining a second loss value based on the first similarity and the second similarity, and training the recommendation prediction model to be trained based on the first loss value and the second loss value.

In the embodiment of the disclosure, a triple loss function is adopted, historical recommended features are used as anchor points (anchors), first-class features are used as positive examples (positive examples), second-class features are used as negative examples (negative examples), and a recommendation prediction model is trained.

In the embodiment of the disclosure, a cross entropy loss mode and a contrast learning training mode are adopted, and the similar examples and the dissimilar examples of the historical recommended features are used for training the recommended prediction model by constructing the first type features and the second type features, so that the distance between the historical recommended features and the similar examples in the feature space can be shortened, the distance between the historical recommended features and the dissimilar examples in the feature space can be further shortened, a self-supervision training mode is realized, and the accuracy of the recommended prediction model is improved.

The second way is: determining a third similarity of the historical search feature and the third type of feature and a fourth similarity of the historical search feature and the fourth type of feature; based on the sample recommendation probability, the prediction recommendation probability, the third similarity and the fourth similarity, training the recommendation prediction model to be trained so that the third similarity is increased and the fourth similarity is reduced, and obtaining the target recommendation prediction model.

In the embodiment of the disclosure, the search interest feature of the sample object is obtained by fusing a resource feature, a historical search feature of sample historical search information, a third type feature and a fourth type feature contained in the historical search feature, wherein the similarity between the third type feature and the historical recommendation feature of the sample historical recommendation information is not less than a second similarity threshold, and the similarity between the fourth type feature and the historical recommendation feature is less than the second similarity threshold.

In the embodiment of the disclosure, the search interest feature of the sample object is obtained by fusing a resource feature based on a sample resource, a historical search feature of sample historical search information, a third type feature and a fourth type feature contained in the historical search feature, wherein the similarity between the third type feature and the historical recommendation feature of the sample historical recommendation information is not less than a second similarity threshold, and the similarity between the fourth type feature and the historical recommendation feature is less than the second similarity threshold.

The third similarity is used for representing the similarity degree of the historical recommended features and the first type of features, and the fourth similarity is used for representing the similarity degree of the historical recommended features and the second type of features. The third similarity and the fourth similarity can be obtained in any manner, for example, in such a manner that the euclidean distance is used to obtain the third similarity and the fourth similarity.

In the embodiment of the disclosure, the historical search feature, the third feature and the fourth feature are all obtained based on a recommendation prediction model, the third feature and the fourth feature are obtained by classifying sub-features in the historical recommendation feature, the similarity between the third feature and the historical recommendation feature is not smaller than a second similarity threshold, the similarity between the fourth feature and the historical recommendation feature is smaller than a second similarity threshold, and the recommendation prediction model to be trained is trained based on the third similarity and the fourth similarity so that the third similarity is increased and the fourth similarity is reduced, that is, the historical search feature obtained based on the recommendation prediction model is more similar to the third feature, and the historical search feature is more dissimilar to the fourth feature, so that the historical search feature obtained based on the recommendation prediction model can reflect the interest of a sample object more, and the accuracy of the recommendation prediction model is improved.

In some embodiments, the second way comprises: and determining a first loss value based on the sample probability and the prediction recommendation probability, determining a third loss value based on the third similarity and the fourth similarity, and training the recommendation prediction model to be trained based on the first loss value and the third loss value.

In the embodiment of the disclosure, a cross entropy loss mode and a contrast learning training mode are adopted, by constructing a third class feature and a fourth class feature as similar examples and dissimilar examples of the historical search feature, the similar examples and the dissimilar examples are utilized to train the recommended prediction model, so that the distance between the historical search feature and the similar example in the feature space is shortened, the distance between the historical search feature and the dissimilar example in the feature space is further shortened, a self-supervision training mode is realized, and the accuracy of the recommended prediction model is improved.

In a third mode, the history search features are formed by fusion features corresponding to each history search text in the sample history search information, and the fusion features corresponding to the history search text are obtained by fusion of text features of the history search text and features of corresponding history search resources: determining a fifth similarity of text features of the historical search text and features of the corresponding historical search resources; and training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity and the fifth similarity so that the third similarity is increased, the fourth similarity is reduced and the fifth similarity is increased, and obtaining the target recommendation prediction model.

In some embodiments, the third way comprises: the method comprises the steps of determining a first loss value based on a sample probability and a prediction recommendation probability, determining a third loss value based on a third similarity and a fourth similarity, determining a fourth loss value based on a fifth similarity, and training a recommendation prediction model based on the first loss value, the third loss value and the fourth loss value.

In the embodiment of the disclosure, considering the similarity between the same historical search text and the corresponding historical search resource, training the recommendation prediction model by utilizing the fifth similarity between the text feature of the historical search text and the feature of the corresponding historical search resource, so that the text feature of the historical search text obtained based on the recommendation prediction model is similar to the feature of the corresponding historical search resource, and further the accuracy of the recommendation prediction model is improved.

In some embodiments, the third way comprises: determining a negative sample resource of a historical search text and a negative sample text of the historical search resource corresponding to the historical search text from the sample historical search information, wherein the negative sample resource is any historical search resource except the historical search resource corresponding to the historical search text in the sample historical search information, and the negative sample text is any historical search text except the historical search text in the sample historical search information; determining a sixth similarity between the historical search text and the negative sample resource and a seventh similarity between the historical search resource corresponding to the historical search text and the negative sample text; training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity, the sixth similarity and the seventh similarity so as to increase the third similarity, decrease the fourth similarity, decrease the sixth similarity and decrease the seventh similarity, and obtaining the target recommendation prediction model.

In the embodiment of the disclosure, the sample history search information can reflect the corresponding relation between the history search text and the history search resource, the history search text with the corresponding relation is similar to the history search resource, and the history search text without the corresponding relation is dissimilar to the history search resource. And taking the history search text with the corresponding relation as a positive sample text of the history search resource, taking the history search text which does not have the corresponding relation with the history search resource in the sample history search information as a negative sample of the history search resource, and similarly, determining positive sample resources and negative sample resources of the history search text.

The sixth similarity is the similarity between the text characteristics of the historical search text obtained based on the recommendation prediction model and the characteristics of the negative sample resource, and the seventh similarity is the similarity between the resource characteristics of the historical search resource obtained based on the recommendation prediction model and the characteristics of the negative sample text. The sixth similarity and the seventh similarity can reflect the accuracy of the features extracted by the recommended prediction model, and further reflect the accuracy of the recommended prediction model.

In some embodiments, the process of training the recommended prediction model includes: based on the sample probability and the predicted recommendation probability, determining a first loss value, based on a third similarity and a fourth similarity, determining a third loss value, based on a fifth similarity, a sixth similarity and a seventh similarity, determining a fourth loss value, and training a recommended prediction model to be trained based on the first loss value, the third loss value and the fourth loss value.

In the embodiment of the disclosure, a contrast learning loss (InfoNCE) can be adopted to align the features of the historical search text and the historical search resource, and a fourth loss value is determined based on the fifth similarity, the sixth similarity and the seventh similarity, so that the recommended prediction model is trained by using the similarity corresponding to the positive sample and the similarity corresponding to the negative sample, the distances between the historical search text and the positive sample resource, between the historical search resource and the positive sample text in the feature space are shortened, and the distances between the historical search text and the negative sample resource, and between the historical search resource and the negative sample text in the feature space are pushed away.

It should be noted that the above description is given by way of example only of any of the three modes, and in another embodiment, the three modes can be arbitrarily combined, for example, the three modes are combined, and the recommended prediction model is trained by using the three modes.

According to the scheme provided by the embodiment of the disclosure, the historical recommendation information and the historical search information are considered to respectively represent which resources are interested by the object in the recommendation scene and the search scene, the historical search information is taken as auxiliary information in the recommendation scene, the historical recommendation information and the historical search information are combined, the interested degree of the resource to be recommended by the object in the recommendation scene and the search scene can be simulated, the sample recommendation probability of the sample object, the sample resource, the sample historical recommendation information of the sample resource and the sample historical search information of the sample object are combined, and the recommendation prediction model is trained, so that the accuracy of the recommendation prediction model is improved, the recommendation probability of the object for executing interactive operation on the resource under the condition that any resource is recommended to any object can be obtained by using the recommendation prediction model after training, and further the resource recommendation effect is ensured.

In addition, in the embodiment of the disclosure, the resource characteristics of the resource are used for representing the resource, the history recommendation characteristics can reflect the history recommendation resource which is interested in the object in the recommendation scene, the first type characteristics accord with the interests of the object in the recommendation scene and the search scene, the second type characteristics accord with the interests of the object in the recommendation scene, and the similarity between the resource characteristics, the history recommendation characteristics, the first type characteristics and the second type characteristics of the resource to be recommended in the recommendation scene and the history recommendation resource which is interested in the object in the recommendation scene is considered, and the historic recommendation resources which are more interested in the object under the influence of the search scene are considered, so that the interest degree of the object to be recommended in the recommendation scene can be simulated, and the accuracy of the determined recommendation interest characteristics is further ensured.

In fig. 5 or fig. 6, the recommended prediction model to be trained is only subjected to one-time iterative training as an example, but in another embodiment, according to the embodiment shown in fig. 6, the recommended prediction model to be trained may be subjected to multiple iterative training, and if the number of iterations reaches a first threshold or the difference between the sample recommendation probability and the prediction recommendation probability in the current iteration round is smaller than a second threshold, the recommended prediction model is stopped being trained, and the recommended prediction model at that time is determined as the target recommended prediction model.

Based on the embodiments shown in fig. 5 to 6, the disclosure further provides a flowchart of a recommended prediction model training method, as shown in fig. 7, where the method includes:

step 1, acquiring a sample object, a sample resource, a sample recommendation probability of the sample resource, sample history recommendation information of the sample object and sample history search information.

Step 2, invoking a recommendation prediction model to be trained, and extracting features of historical recommendation resources in the sample historical recommendation information to obtain a recommendation resource feature matrix, wherein the recommendation feature matrix comprises resource features of each historical recommendation resource; extracting features of historical search texts in the sample historical search information to obtain a text feature matrix, wherein the text feature matrix comprises text features of each historical search text; and extracting features of the historical search resources in the sample historical search information to obtain a search resource feature matrix, wherein the search feature matrix comprises resource features of each historical search resource.

And step 3, calling a recommended prediction model to be trained, and fusing each resource feature in the recommended resource feature matrix with the corresponding first position feature by adopting deviation coding to obtain an updated recommended resource feature matrix.

And 4, calling a recommendation prediction model to be trained, determining the corresponding resource characteristics of each text characteristic in the text characteristic matrix in the search resource characteristic matrix based on the historical search resources corresponding to each historical search text, and fusing each text characteristic in the text characteristic matrix with the corresponding resource characteristic, the second position characteristic corresponding to the resource characteristic and the search type characteristic by adopting deviation coding to obtain an updated text characteristic matrix.

Step 5, calling a first attention sub-model in a recommended prediction model to be trained, and updating each sub-feature based on a plurality of sub-features in the updated recommended resource feature matrix to obtain a recommended feature matrix; and calling a second attention sub-model in the recommended prediction model, and updating each sub-feature based on a plurality of sub-features in the updated text feature matrix to obtain a search feature matrix.

The recommended feature matrix is the historical recommended feature in the above embodiment, and the search feature matrix is the historical search feature in the above embodiment. The first attention sub-model and the second attention sub-model are arbitrary network models, for example, the first attention sub-model and the second attention sub-model are both a transducer (one network model), an RNN (Recurrent Neural Network ) model, a GRU unit (sequential recommendation model), and the like, but model parameters in the first attention sub-model and the second attention sub-model are different.

Step 6, calling a common attention sub model in a recommended prediction model to be trained, and comparing the recommended feature matrix with the search feature matrix to obtain first similarity information and second similarity information; dividing the recommended feature matrix based on the first similarity information to obtain a first type of features and a second type of features contained in the recommended feature matrix; and dividing the search feature matrix based on the second similarity information to obtain a third type of features and a fourth type of features contained in the search feature matrix.

The common Attention sub-model is any network model, for example, co-Attention.

And 7, calling a multi-head attention sub-model in the recommendation prediction model to be trained, adopting a multi-head attention mechanism, respectively fusing the historical recommendation feature, the first type feature and the second type feature with the resource feature of the resource to obtain the similarity between the resource feature of the resource and the historical recommendation feature, the first type feature and the second type feature, determining the product of the historical recommendation feature and the corresponding similarity, the product of the first type feature and the corresponding similarity and the product of the second type feature and the corresponding similarity, and splicing the product of the historical recommendation feature and the corresponding similarity, the product of the first type feature and the corresponding similarity and the product of the second type feature and the corresponding similarity to obtain the recommendation interest feature matrix.

The Multi-head attention sub-model is an arbitrary network model, for example, multi-head attention sub-model is Multi-interest Extraction (Multi-head interest extraction model).

And 8, calling a multi-head attention sub-model in the recommendation prediction model, adopting a multi-head attention mechanism, respectively fusing the historical search feature, the third type feature and the fourth type feature with the resource feature of the resource to obtain the similarity between the resource feature of the resource and the historical search feature, the third type feature and the fourth type feature, determining the product of the historical search feature and the corresponding similarity, the product of the third type feature and the corresponding similarity and the product of the fourth type feature and the corresponding similarity, and splicing the product of the historical search feature and the corresponding similarity, the product of the third type feature and the corresponding similarity and the product of the fourth type feature and the corresponding similarity to obtain the search interest feature matrix.

Step 9, calling a recommended prediction model to be trained, and splicing object features of the sample object, resource features of sample resources, recommended interest feature matrixes and search interest feature matrixes to obtain sample splicing features; and calling a feature conversion sub-model in the recommendation prediction model, and performing dimension reduction processing on the sample splicing features to obtain the prediction recommendation probability of the sample resources.

Wherein the feature transformation sub-model is an arbitrary network model, for example, the feature transformation sub-model is a multi-layer perceptron (MLP, multilayer Perceptron)

Step 10, determining negative sample resources of the historical search text and negative sample texts of the historical search resources corresponding to the historical search text from the sample historical search information; invoking a recommended prediction model to be trained, and acquiring the characteristics of the negative sample resources and the characteristics of the negative sample text; determining a fifth similarity of text features of the historical search text and features of the corresponding historical search resources; determining a sixth similarity between the historical search text and the negative sample resource, and a seventh similarity between the historical search resource corresponding to the historical search text and the negative sample text, and determining a fourth loss value based on the fifth similarity, the sixth similarity and the seventh similarity; determining a first similarity of the historical recommended features and the first type of features and a second similarity of the historical recommended features and the second type of features, and determining a second loss value based on the first similarity and the second similarity; determining a third similarity of the historical search feature and the third type of feature, and a fourth similarity of the historical search feature and the third type of feature, and determining a third loss value based on the third similarity and the fourth similarity; determining a first loss value based on the sample probability and the predicted recommendation probability; and training the recommended prediction model to be trained based on the first loss value, the second loss value, the third loss value and the fourth loss value to obtain a target recommended prediction model.

According to the recommendation prediction model training method provided by the embodiment, after the recommendation prediction model is trained, the performance of the obtained target recommendation prediction model is improved. For example, for the performance index NDCG (Normalized Discounted Cumulative Gain, normalized break cumulative gain), ndcg@5 is raised by 4.97% in the case of considering the first 5 of the test resources, and ndcg@10 is raised by 7.05% in the case of considering the first 10 of the test resources; for the performance index HIT (HIT rate), hit@1 is improved by 6.20% with the first 1 resources in the test resources considered; hit@5 is raised by 4.23% taking into account the first 5 of the test resources; taking into account the first 10 of the test resources, hit@10 is raised by 7.48% and MRR (Mean Reciprocal Rank, average reciprocal rank) is raised by 6.49%.

It should be noted that, all the above technical solutions can be combined arbitrarily to form an optional embodiment of the disclosure, which is not described herein in detail.

FIG. 8 is a block diagram of a resource recommendation device, as shown in FIG. 8, according to an exemplary embodiment, the device comprising:

an acquisition unit 801 configured to perform acquisition of history recommendation information of an object, the history recommendation information indicating history recommendation resources of the object having performed an interactive operation, and history search information indicating history search text input by the object and history search resources searched based on the history search text and the object having performed an interactive operation;

The obtaining unit 801 is further configured to perform obtaining a recommended interest feature and a search interest feature of the object based on the resource to be recommended, the history recommendation information, and the history search information, where the recommended interest feature indicates an interest degree of the object in the resource in the recommendation scene, and the search interest feature indicates an interest degree of the object in the resource in the search scene;

the processing unit 802 is configured to perform dimension reduction processing on object features of the object, resource features of the resource, recommended interest features and search interest features to obtain recommended probability of the resource, where the recommended probability indicates a possibility that the object performs an interactive operation on the resource when recommending the resource to the object;

and a recommendation unit 803 configured to perform recommendation of a target resource among the plurality of resources to the object based on recommendation probabilities of the plurality of resources to be recommended.

In some embodiments, the obtaining unit 801 is configured to perform feature extraction on the historical recommendation information and the historical search information respectively to obtain a historical recommendation feature and a historical search feature, where the historical recommendation feature includes a sub-feature for characterizing a historical recommendation resource, and the historical search feature includes a sub-feature for characterizing a historical search text and a historical search resource corresponding to the historical search text; classifying sub-features in the history recommended features and sub-features in the history search features respectively to obtain a first type of features and a second type of features contained in the history recommended features and a third type of features and a fourth type of features contained in the history search features, wherein the similarity between the first type of features and the history search features is not smaller than a first similarity threshold, the similarity between the second type of features and the history search features is smaller than the first similarity threshold, the similarity between the third type of features and the history recommended features is not smaller than a second similarity threshold, and the similarity between the fourth type of features and the history recommended features is smaller than the second similarity threshold; based on the similarity between the resource characteristics of the resources and the historical recommendation characteristics, the first type characteristics and the second type characteristics, the historical recommendation characteristics, the first type characteristics and the second type characteristics are fused to obtain recommendation interest characteristics; and fusing the historical search features, the third type features and the fourth type features based on the similarity between the resource features of the resource and the historical search features, the third type features and the fourth type features respectively to obtain search interest features.

In some embodiments, the obtaining unit 801 is configured to perform feature extraction on each historical recommended resource in the historical recommendation information to obtain a resource feature of each historical recommended resource, and form the resource feature of each historical recommended resource into a historical recommended feature; respectively extracting features of the historical search text and the historical search resources in the historical search information to obtain text features of each historical search text and resource features of each historical search resource; and fusing the text characteristics of each historical search text and the resource characteristics of the corresponding historical search resource to obtain first fusion characteristics corresponding to each historical search text, and forming the first fusion characteristics corresponding to each historical search text into the historical search characteristics.

In some embodiments, the obtaining unit 801 is configured to perform obtaining a first location feature of each historical recommended resource, the first location feature indicating a relative time order between the historical recommended resource and other historical recommended resources in the historical recommendation information; fusing the resource characteristics of each historical recommended resource with the first position characteristics to obtain fusion characteristics corresponding to each historical recommended resource; based on the fusion characteristics corresponding to the historical recommended resources, updating the fusion characteristics corresponding to each historical recommended resource respectively, and forming the historical recommended characteristics by updating the characteristics of the historical recommended resources.

In some embodiments, the obtaining unit 801 is configured to perform obtaining a second location feature of each history search text, and a search type feature, where the second location feature indicates a relative time order between the history search text and other history search texts in the history search information, and the search type feature indicates a search type employed when searching based on the history search text; fusing the first fusion feature, the second position feature and the search type feature corresponding to each historical search text to obtain a second fusion feature corresponding to each historical search text; and updating the second fusion features corresponding to each historical search text based on the second fusion features corresponding to the plurality of historical search texts, and forming the historical search features by the updated features of the plurality of historical search texts.

In some embodiments, the obtaining unit 801 is configured to perform comparing the historical recommended feature with the historical search feature to obtain first similarity information and second similarity information, where the first similarity information indicates a similarity between each sub-feature in the historical recommended feature and the historical search feature, and the second similarity information indicates a similarity between each sub-feature in the historical search feature and the historical recommended feature; classifying sub-features in the historical recommended features based on the first similarity information to obtain first-class features and second-class features; and classifying the sub-features in the historical search features based on the second similarity information to obtain a third type of features and a fourth type of features.

In some embodiments, the processing unit 802 is configured to perform stitching on the object feature of the object, the resource feature of the resource, the recommended interest feature, and the search interest feature to obtain a stitched feature; and performing dimension reduction processing on the spliced features to obtain recommendation probability.

It should be noted that, in the apparatus provided in the foregoing embodiment, only the division of the functional units is illustrated, and in practical application, the functional allocation may be performed by different functional units according to needs, that is, the internal structure of the electronic device is divided into different functional units, so as to perform all or part of the functions described above. In addition, the recommended prediction model training device and the recommended prediction model training method provided in the foregoing embodiments belong to the same concept, and detailed implementation processes of the recommended prediction model training device and the recommended prediction model training method are detailed in the method embodiments, which are not repeated here.

The specific manner in which the various modules perform the operations in the apparatus of the above embodiments have been described in detail in connection with the embodiments of the method, and will not be described in detail herein.

FIG. 9 is a block diagram of a recommendation prediction model training apparatus, as shown in FIG. 9, according to an exemplary embodiment, comprising:

An acquisition unit 901 configured to perform acquisition of a sample object, a sample resource, a sample recommendation probability of the sample resource, sample history recommendation information of the sample object, and sample history search information indicating a history recommendation resource of the sample object on which an interactive operation was performed, the sample history search information indicating a history search text input by the sample object and a history search resource searched based on the history search text and on which the sample object was performed;

the obtaining unit 901 is further configured to execute invoking a recommendation prediction model to be trained, obtain a recommendation interest feature of the sample object and a search interest feature based on the sample resource, the sample history recommendation information and the sample history search information, wherein the recommendation interest feature indicates an interest degree of the sample object to the sample resource in a recommendation scene, and the search interest feature indicates an interest degree of the sample object to the sample resource in a search scene;

the processing unit 902 is configured to execute invoking a recommendation prediction model to be trained, perform dimension reduction processing on object features of the sample object, resource features of the sample resource, recommendation interest features and search interest features, and obtain a prediction recommendation probability of the sample resource, where the prediction recommendation probability indicates a possibility that the sample object performs an interactive operation on the sample resource when recommending the sample resource to the sample object;

The training unit 903 is configured to perform training on the recommended prediction model to be trained based on the sample recommendation probability and the predicted recommendation probability, so as to obtain a target recommended prediction model.

In some embodiments, the recommended interest feature of the sample object is obtained by fusing a resource feature, a historical recommendation feature of sample historical recommendation information, a first type feature and a second type feature contained in the historical recommendation feature, wherein the similarity between the first type feature and the historical search feature of the sample historical search information is not less than a first similarity threshold, the similarity between the second type feature and the historical search feature is less than the first similarity threshold, the first type feature and the second type feature are obtained by classifying sub-features in the historical recommendation feature, and the sub-features in the historical recommendation feature are used for representing historical recommendation resources; as shown in fig. 10, the apparatus further includes:

a determining unit 904 configured to perform determining a first similarity of the history recommended feature to the first type of feature and a second similarity of the history recommended feature to the second type of feature;

the training unit 903 is configured to perform training on the recommended prediction model to be trained based on the sample recommendation probability, the predicted recommendation probability, the first similarity, and the second similarity, so that the first similarity increases and the second similarity decreases, and obtain the target recommended prediction model.

In some embodiments, the search interest feature of the sample object is obtained by fusing a resource feature, a historical search feature of sample historical search information, a third type feature and a fourth type feature contained in the historical search feature, the similarity between the third type feature and the historical recommendation feature of the sample historical recommendation information is not less than a second similarity threshold, the similarity between the fourth type feature and the historical recommendation feature is less than the second similarity threshold, the third type feature and the fourth type feature are obtained by classifying sub-features in the historical search feature, and the sub-features in the historical search feature are used for representing historical search texts and historical search resources corresponding to the historical search texts; as shown in fig. 10, the apparatus further includes:

a determining unit 904 configured to perform determining a third similarity of the history search feature to the third class of features, a fourth similarity of the history search feature to the fourth class of features;

the training unit 903 is configured to perform training on the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, and the fourth similarity, so that the third similarity increases and the fourth similarity decreases, and obtain the target recommendation prediction model.

In some embodiments, the history search features are formed by fusion features corresponding to each history search text in the sample history search information, and the fusion features corresponding to the history search text are obtained by fusion of text features of the history search text and features of corresponding history search resources; as shown in fig. 10, the apparatus further includes:

a determining unit 904 configured to perform determining a fifth similarity of text features of the history search text with features of the corresponding history search resource;

the training unit 903 is configured to perform training on the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity, and the fifth similarity, so that the third similarity increases, the fourth similarity decreases, and the fifth similarity increases, and obtain the target recommendation prediction model.

In some embodiments, as shown in fig. 10, the apparatus further comprises:

a determining unit 904 configured to determine, from the sample history search information, a negative sample resource of the history search text, a negative sample text of the history search resource corresponding to the history search text, the negative sample resource being any history search resource other than the history search resource corresponding to the history search text in the sample history search information, the negative sample text being any history search text other than the history search text in the sample history search information;

A determining unit 904 further configured to perform determining a sixth similarity of the history search text to the negative sample text, and a seventh similarity of the history search resource corresponding to the history search text to the negative sample text;

the training unit 903 is configured to perform training on the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity, the sixth similarity, and the seventh similarity, so as to increase the third similarity, decrease the fourth similarity, decrease the sixth similarity, and decrease the seventh similarity, thereby obtaining the target recommendation prediction model.

When the electronic device is provided as a terminal, fig. 11 is a block diagram of a terminal 1100 according to an exemplary embodiment. Fig. 11 shows a block diagram of a terminal 1100 according to an exemplary embodiment of the present disclosure. Generally, the terminal 1100 includes: a processor 1101 and a memory 1102.

The processor 1101 may include one or more processing cores, such as a 4-core processor, an 8-core processor, and the like. The processor 1101 may be implemented in at least one hardware form of DSP (Digital Signal Processing ), FPGA (Field-Programmable Gate Array, field programmable gate array), PLA (Programmable Logic Array ). The processor 1101 may also include a main processor, which is a processor for processing data in an awake state, also called a CPU (Central Processing Unit ), and a coprocessor; a coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 1101 may integrate a GPU (Graphics Processing Unit, image processor) for rendering and drawing of content required to be displayed by the display screen. In some embodiments, the processor 1101 may also include an AI (Artificial Intelligence ) processor for processing computing operations related to machine learning.

Memory 1102 may include one or more computer-readable storage media, which may be non-transitory. Memory 1102 may also include high-speed random access memory, as well as non-volatile memory, such as one or more magnetic disk storage devices, flash memory storage devices. In some embodiments, a non-transitory computer readable storage medium in memory 1102 is used to store at least one program code for execution by processor 1101 to implement the resource recommendation method or recommendation prediction model training method provided by the method embodiments in the present disclosure.

In some embodiments, the terminal 1100 may further optionally include: a peripheral interface 1103 and at least one peripheral. The processor 1101, memory 1102, and peripheral interface 1103 may be connected by a bus or signal lines. The individual peripheral devices may be connected to the peripheral device interface 1103 by buses, signal lines or circuit boards. Specifically, the peripheral device includes: at least one of radio frequency circuitry 1104, a display screen 1105, a camera assembly 1106, audio circuitry 1107, and a power supply 1108.

A peripheral interface 1103 may be used to connect I/O (Input/Output) related at least one peripheral device to the processor 1101 and memory 1102. In some embodiments, the processor 1101, memory 1102, and peripheral interface 1103 are integrated on the same chip or circuit board; in some other embodiments, any one or both of the processor 1101, memory 1102, and peripheral interface 1103 may be implemented on a separate chip or circuit board, which is not limited in this embodiment.

The Radio Frequency circuit 1104 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency circuit 1104 communicates with a communication network and other communication devices via electromagnetic signals. The radio frequency circuit 1104 converts an electrical signal into an electromagnetic signal for transmission, or converts a received electromagnetic signal into an electrical signal. Optionally, the radio frequency circuit 1104 includes: antenna systems, RF transceivers, one or more amplifiers, tuners, oscillators, digital signal processors, codec chipsets, subscriber identity module cards, and so forth. The radio frequency circuitry 1104 may communicate with other terminals via at least one wireless communication protocol. The wireless communication protocol includes, but is not limited to: metropolitan area networks, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and/or WiFi (Wireless Fidelity ) networks. In some embodiments, the radio frequency circuitry 1104 may also include NFC (Near Field Communication, short range wireless communication) related circuitry, which is not limited by this disclosure.

The display screen 1105 is used to display a UI (User Interface). The UI may include graphics, text, icons, video, and any combination thereof. When the display 1105 is a touch display, the display 1105 also has the ability to collect touch signals at or above the surface of the display 1105. The touch signal may be input to the processor 1101 as a control signal for processing. At this time, the display screen 1105 may also be used to provide virtual buttons and/or virtual keyboards, also referred to as soft buttons and/or soft keyboards. In some embodiments, the display 1105 may be one, providing a front panel of the terminal 1100; in other embodiments, the display 1105 may be at least two, respectively disposed on different surfaces of the terminal 1100 or in a folded design; in still other embodiments, the display 1105 may be a flexible display disposed on a curved surface or a folded surface of the terminal 1100. Even more, the display 1105 may be arranged in a non-rectangular irregular pattern, i.e., a shaped screen. The display 1105 may be made of LCD (Liquid Crystal Display ), OLED (Organic Light-Emitting Diode) or other materials.

The camera assembly 1106 is used to capture images or video. Optionally, the camera assembly 1106 includes a front camera and a rear camera. Typically, the front camera is disposed on the front panel of the terminal and the rear camera is disposed on the rear surface of the terminal. In some embodiments, the at least two rear cameras are any one of a main camera, a depth camera, a wide-angle camera and a tele camera, so as to realize that the main camera and the depth camera are fused to realize a background blurring function, and the main camera and the wide-angle camera are fused to realize a panoramic shooting and Virtual Reality (VR) shooting function or other fusion shooting functions. In some embodiments, the camera assembly 1106 may also include a flash. The flash lamp can be a single-color temperature flash lamp or a double-color temperature flash lamp. The dual-color temperature flash lamp refers to a combination of a warm light flash lamp and a cold light flash lamp, and can be used for light compensation under different color temperatures.

The audio circuit 1107 may include a microphone and a speaker. The microphone is used for collecting sound waves of users and environments, converting the sound waves into electric signals, and inputting the electric signals to the processor 1101 for processing, or inputting the electric signals to the radio frequency circuit 1104 for voice communication. For purposes of stereo acquisition or noise reduction, a plurality of microphones may be provided at different portions of the terminal 1100, respectively. The microphone may also be an array microphone or an omni-directional pickup microphone. The speaker is used to convert electrical signals from the processor 1101 or the radio frequency circuit 1104 into sound waves. The speaker may be a conventional thin film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, not only the electric signal can be converted into a sound wave audible to humans, but also the electric signal can be converted into a sound wave inaudible to humans for ranging and other purposes. In some embodiments, the audio circuit 1107 may also include a headphone jack.

A power supply 1108 is used to power the various components in terminal 1100. The power supply 1108 may be an alternating current, a direct current, a disposable battery, or a rechargeable battery. When the power supply 1108 includes a rechargeable battery, the rechargeable battery may support wired or wireless charging. The rechargeable battery may also be used to support fast charge technology.

Those skilled in the art will appreciate that the structure shown in fig. 11 is not limiting and that terminal 1100 may include more or fewer components than shown, or may combine certain components, or may employ a different arrangement of components.

When the electronic device is provided as a server, fig. 12 is a block diagram of a server 1200 according to an exemplary embodiment, where the server 1200 may be greatly different due to configuration or performance, and may include one or more processors (Central Processing Units, CPU) 1201 and one or more memories 1202, where at least one program code is stored in the memories 1202, and the at least one program code is loaded and executed by the processor 1201 to implement the resource recommendation method or the recommendation prediction model training method provided in the above method embodiments. Of course, the server may also have a wired or wireless network interface, a keyboard, an input/output interface, etc. to perform input/output, and the server 1200 may also include other components for implementing the functions of the device, which are not described herein.

In an exemplary embodiment, a computer readable storage medium is also provided, which when executed by a processor of an electronic device, causes the electronic device to perform the resource recommendation method or the recommendation prediction model training method in the above-described embodiments.

In an exemplary embodiment, a computer program product is also provided, comprising a computer program/instruction which, when executed by a processor, implements the resource recommendation method or recommendation prediction model training method in the above-described embodiments.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This disclosure is intended to cover any adaptations, uses, or adaptations of the disclosure following the general principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It is to be understood that the present disclosure is not limited to the precise arrangements and instrumentalities shown in the drawings, and that various modifications and changes may be effected without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for recommending resources, the method comprising:

2. The method of claim 1, wherein the obtaining recommended interest features and search interest features of the object based on the resources to be recommended, the historical recommendation information, and the historical search information comprises:

respectively extracting features of the historical recommendation information and the historical search information to obtain historical recommendation features and historical search features, wherein the historical recommendation features comprise sub-features used for representing the historical recommendation resources, and the historical search features comprise sub-features used for representing the historical search texts and the historical search resources corresponding to the historical search texts;

Based on the similarity between the resource characteristics of the resources and the historical recommendation characteristics, the first type characteristics and the second type characteristics, the historical recommendation characteristics, the first type characteristics and the second type characteristics are fused to obtain the recommendation interest characteristics;

and fusing the historical search feature, the third type feature and the fourth type feature based on the similarity between the resource feature of the resource and the historical search feature, the third type feature and the fourth type feature respectively to obtain the search interest feature.

3. The method of claim 2, wherein the feature extracting the historical recommendation information and the historical search information to obtain the historical recommendation feature and the historical search feature respectively comprises:

extracting features of each historical recommended resource in the historical recommended information to obtain resource features of each historical recommended resource, and forming the resource features of each historical recommended resource into the historical recommended features;

extracting features of the historical search text and the historical search resources in the historical search information respectively to obtain text features of each historical search text and resource features of each historical search resource;

4. A method according to claim 3, wherein said composing the resource characteristics of each of the history recommended resources into the history recommended characteristics comprises:

acquiring a first position characteristic of each historical recommended resource, wherein the first position characteristic indicates a relative time sequence between the historical recommended resource and other historical recommended resources in the historical recommended information;

based on the fusion characteristics corresponding to the historical recommended resources, updating the fusion characteristics corresponding to each historical recommended resource respectively, and forming the historical recommended characteristics by the updated characteristics of the historical recommended resources.

5. The method of claim 3, wherein said constructing the first fusion feature corresponding to each of the historical search text into the historical search feature comprises:

6. The method of claim 2, wherein classifying the sub-features in the historical recommended features and the sub-features in the historical search features to obtain the first class of features and the second class of features included in the historical recommended features and the third class of features and the fourth class of features included in the historical search features, respectively, comprises:

Comparing the historical recommended features with the historical search features to obtain first similarity information and second similarity information, wherein the first similarity information indicates the similarity between each sub-feature in the historical recommended features and the historical search features, and the second similarity information indicates the similarity between each sub-feature in the historical search features and the historical recommended features;

classifying sub-features in the historical recommended features based on the first similarity information to obtain the first type features and the second type features;

and classifying the sub-features in the historical search features based on the second similarity information to obtain the third type of features and the fourth type of features.

7. The method according to any one of claims 1-6, wherein performing the dimension reduction processing on the object feature of the object, the resource feature of the resource, the recommended interest feature, and the search interest feature to obtain the recommended probability of the resource includes:

splicing the object features of the object, the resource features of the resource, the recommended interest features and the search interest features to obtain splicing features;

And performing dimension reduction processing on the spliced features to obtain the recommendation probability.

8. A recommended predictive model training method, the method further comprising:

9. The method of claim 8, wherein the recommended interest feature of the sample object is obtained by fusing a resource feature, a historical recommendation feature of the sample historical recommendation information, a first type feature and a second type feature included in the historical recommendation feature, wherein a similarity between the first type feature and a historical search feature of the sample historical search information is not less than a first similarity threshold, a similarity between the second type feature and the historical search feature is less than a first similarity threshold, and the first type feature and the second type feature are obtained by classifying sub-features in the historical recommendation feature, wherein the sub-features in the historical recommendation feature are used for representing the historical recommendation resource; the method further comprises the steps of training the recommended prediction model to be trained based on the sample recommendation probability and the predicted recommendation probability, and before obtaining a target recommended prediction model:

determining a first similarity of the historical recommendation feature and the first type of feature and a second similarity of the historical recommendation feature and the second type of feature;

Training the recommended prediction model to be trained based on the sample recommendation probability and the predicted recommendation probability to obtain a target recommended prediction model, including:

and training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the first similarity and the second similarity so as to increase the first similarity and decrease the second similarity, thereby obtaining the target recommendation prediction model.

10. The method according to claim 8, wherein the search interest feature of the sample object is obtained by fusing the resource feature, the history search feature of the sample history search information, and a third class feature and a fourth class feature included in the history search feature, a similarity between the third class feature and the history recommendation feature of the sample history recommendation information is not less than a second similarity threshold, a similarity between the fourth class feature and the history recommendation feature is less than a second similarity threshold, and the third class feature and the fourth class feature are obtained by classifying sub-features in the history search feature, wherein the sub-features in the history search feature are used for characterizing the history search text and the history search resource corresponding to the history search text; the method further comprises the steps of training the recommended prediction model to be trained based on the sample recommendation probability and the predicted recommendation probability, and before obtaining a target recommended prediction model:

Determining a third similarity of the historical search feature to the third class of features, a fourth similarity of the historical search feature to the fourth class of features;

and training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity and the fourth similarity so as to increase the third similarity and decrease the fourth similarity, thereby obtaining the target recommendation prediction model.

11. The method according to claim 10, wherein the history search feature is formed by a fusion feature corresponding to each history search text in the sample history search information, and the fusion feature corresponding to the history search text is obtained by fusing a text feature of the history search text and a feature of a corresponding history search resource; before the training of the recommended prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity and the fourth similarity, the method further includes:

Determining a fifth similarity between the text features of the historical search text and the features of the corresponding historical search resources;

the training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity and the fourth similarity, so that the third similarity is increased and the fourth similarity is decreased, to obtain the target recommendation prediction model, includes:

and training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity and the fifth similarity so as to increase the third similarity, decrease the fourth similarity and increase the fifth similarity, thereby obtaining the target recommendation prediction model.

12. The method of claim 10, wherein the method further comprises, prior to training the recommended prediction model to be trained based on the sample recommendation probability, the predicted recommendation probability, the third similarity, and the fourth similarity:

determining a negative sample resource of the historical search text and a negative sample text of the historical search resource corresponding to the historical search text from the sample historical search information, wherein the negative sample resource is any historical search resource except the historical search resource corresponding to the historical search text in the sample historical search information, and the negative sample text is any historical search text except the historical search text in the sample historical search information;

training the recommendation prediction model to be trained based on the sample recommendation probability, the prediction recommendation probability, the third similarity, the fourth similarity, the sixth similarity and the seventh similarity, so that the third similarity is increased, the fourth similarity is reduced, the sixth similarity and the seventh similarity are reduced, and the target recommendation prediction model is obtained.

13. A resource recommendation device, the device comprising:

14. A recommended prediction model training apparatus, the apparatus further comprising:

15. An electronic device, the electronic device comprising:

one or more processors;

a memory for storing the processor-executable program code;

wherein the processor is configured to execute the program code to implement the resource recommendation method of any of claims 1 to 7; alternatively, to implement the recommended prediction model training method of any of claims 8 to 12.

16. A computer readable storage medium, characterized in that instructions in the computer readable storage medium, when executed by a processor of an electronic device, enable the electronic device to perform the resource recommendation method of any one of claims 1 to 7; alternatively, to implement the recommended prediction model training method of any of claims 8 to 12.