CN114662002A

CN114662002A - Object recommendation method, medium, device and computing equipment

Info

Publication number: CN114662002A
Application number: CN202210361988.9A
Authority: CN
Inventors: 林浩; 苗壮; 刘森茂; 杨杨
Original assignee: Hangzhou Netease Cloud Music Technology Co Ltd
Current assignee: Hangzhou Netease Cloud Music Technology Co Ltd
Priority date: 2022-04-07
Filing date: 2022-04-07
Publication date: 2022-06-24

Abstract

The embodiment of the disclosure provides an object recommendation method, medium, device and computing equipment. The method comprises the following steps: responding to the received first search request, and acquiring a first object for pushing; generating a second search request in response to the interactive operation on the first object; acquiring a second object to be pushed based on the second search request; and determining a target object for pushing in the second object based on the second search request, the second object and the user characteristic and the object scoring model corresponding to the second search request. The pushed target object is determined again according to the first search request and the first object subjected to interactive operation, so that the target object can reflect the interest tendency expressed by clicking the pushed object by the user, and the user experience is improved.

Description

Object recommendation method, medium, device and computing equipment

Technical Field

Embodiments of the present disclosure relate to the field of computer technologies, and in particular, to an object recommendation method, medium, apparatus, and computing device.

Background

This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims. The description herein is not admitted to be prior art by inclusion in this section.

With the development of the internet, a large amount of video and audio contents appear, and users are used to find interesting contents through a search technology. The common search mode is that the user returns to the object list after inputting the search word, the user clicks the object list to enter the detail page of the pushed object to browse the object content, and browses the content of the next object through sliding operation or other switching operation.

In the existing search technology, the next object usually corresponds to the next object in an object list obtained based on a search word, but the obtained next object usually cannot reflect the interest tendency expressed by a user through clicking a push object, so that the user experience is insufficient.

Disclosure of Invention

The present disclosure provides an object recommendation method, medium, device, and computing device, to solve a problem in the prior art that when a next object is selected based on a push object selected in an object list, the next object cannot reflect an interest tendency expressed by a user by clicking the push object, resulting in insufficient user experience.

In a first aspect of embodiments of the present disclosure, there is provided an object recommendation method, including:

responding to the received first search request, and acquiring first objects for pushing, wherein at least one first object is provided;

responding to the interactive operation of the first object, and generating a second search request, wherein the second search request is generated according to the first search request and the label of the first object corresponding to the interactive operation;

acquiring second objects to be pushed based on the second search request, wherein at least one second object is acquired;

and determining a target object for pushing in the second object based on the second search request, the second object and the user characteristics and the object scoring model corresponding to the second search request.

In a second aspect of the disclosed embodiments, there is provided a method for training a score model of an object, comprising:

obtaining a data set to be extracted based on a client log and a service end log in a server, wherein the category of the data set to be extracted comprises a second search request characteristic, a user characteristic, an object characteristic and a user interaction operation record, the second search request is generated according to the first search request and a label of a first object of the interaction operation, and the first object is obtained according to the first search request in a pushing mode;

extracting corresponding data characteristics according to the category of the data set to be extracted;

processing the data characteristics to obtain training data;

the training data is input into the object scoring model, and the object scoring model is trained based on the loss function.

In a third aspect of embodiments of the present disclosure, there is provided a computer-readable storage medium comprising:

the computer-readable storage medium has stored therein computer-executable instructions for implementing the method as in the first aspect of the present disclosure and/or the second aspect of the present disclosure when executed by the processor.

In a fourth aspect of embodiments of the present disclosure, there is provided an object recommendation apparatus including:

the first acquisition module is used for responding to the received first search request and acquiring first objects for pushing, wherein at least one first object is obtained;

the generating module is used for responding to the interactive operation of the first object and generating a second search request, and the second search request is generated according to the first search request and the label of the first object corresponding to the interactive operation;

the second acquisition module is used for acquiring second objects to be pushed based on a second search request, and at least one second object is acquired;

and the determining module is used for determining a target object for pushing in the second object based on the second search request characteristic, the second object characteristic and the user characteristic and the object scoring model corresponding to the second search request.

In a fifth aspect of the disclosed embodiments, there is provided an object scoring model training apparatus, comprising:

the acquisition module is used for acquiring a data set to be extracted and a labeled data set based on a client log and a server log in a server, the category of the data set to be extracted comprises a second search request characteristic, a user characteristic and a second object characteristic, the labeled data set comprises labeled data of user interaction operation, the second search request is generated according to a first search request and a label of a first object corresponding to the interaction operation, the first object is obtained by pushing according to the first search request, and the second object is obtained according to the second search request;

the extraction module is used for extracting corresponding data characteristics according to the category of the data set to be extracted;

the processing module is used for inputting the preprocessed data characteristics into the object scoring model to obtain a prediction score;

and the training module is used for calculating a loss function based on the prediction score and the labeled data and training the object score model based on the loss function.

In a sixth aspect of embodiments of the present disclosure, there is provided a computing device comprising: at least one processor;

and a memory communicatively coupled to the at least one processor;

wherein the memory stores instructions executable by the at least one processor to cause the computing device to perform the method as in the first aspect of the disclosure and/or the second aspect of the disclosure.

According to the object recommendation method, the medium, the device and the computing equipment, the first object used for pushing is obtained by responding to the received first search request; generating a second search request in response to the interactive operation on the first object; acquiring a second object to be pushed based on the second search request; and determining a target object for pushing in the second object based on the second search request, the second object and the user characteristic and the object scoring model corresponding to the second search request. Therefore, the pushed target object can be determined again according to the first search request and the first object subjected to interactive operation, so that the target object can reflect the interest tendency expressed by interactive operation and the retrieval tendency expressed by the first search request at the same time, the requirements of users can be better met, and the user experience is improved.

Drawings

The above and other objects, features and advantages of exemplary embodiments of the present disclosure will become readily apparent from the following detailed description read in conjunction with the accompanying drawings. Several embodiments of the present disclosure are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings and in which:

FIG. 1 schematically illustrates an application scenario diagram according to an embodiment of the present disclosure;

FIG. 2 schematically shows a flow chart of an object recommendation method according to another embodiment of the present disclosure;

FIG. 3 schematically shows a flow chart of an object recommendation method according to yet another embodiment of the present disclosure;

FIG. 4a is a schematic structural diagram of an object scoring model in the embodiment of FIG. 3;

FIG. 4b schematically shows a flowchart of calculating a second object score by the object scoring model in the embodiment of FIG. 3;

FIG. 5 schematically illustrates a flow chart of an object recommendation method according to yet another embodiment of the present disclosure;

FIG. 6 is a flow chart schematically illustrating the establishment of the tag database in the embodiment of FIG. 5;

FIG. 7 schematically shows a flow chart of a method of object scoring model training in accordance with a further embodiment of the present disclosure;

fig. 8a schematically shows a structural diagram of an object scoring model according to a further embodiment of the present disclosure;

FIG. 8b schematically shows a flow chart of a method of object scoring model training according to a further embodiment of the present disclosure;

FIG. 9 schematically illustrates a structural diagram of a storage medium according to yet another embodiment of the present disclosure;

fig. 10 schematically shows a structural diagram of an object recommending apparatus according to still another embodiment of the present disclosure;

fig. 11 schematically shows a structural diagram of an object scoring model training apparatus according to still another embodiment of the present disclosure;

fig. 12 schematically shows a structural diagram of a computing device according to yet another embodiment of the present disclosure.

In the drawings, like or corresponding reference characters designate like or corresponding parts.

Detailed Description

The principles and spirit of the present disclosure will be described with reference to a number of exemplary embodiments. It is understood that these embodiments are given solely for the purpose of enabling those skilled in the art to better understand and to practice the present disclosure, and are not intended to limit the scope of the present disclosure in any way. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

As will be appreciated by one skilled in the art, embodiments of the present disclosure may be embodied as a system, apparatus, device, method, or computer program product. Accordingly, the present disclosure may be embodied in the form of: entirely hardware, entirely software (including firmware, resident software, micro-code, etc.), or a combination of hardware and software.

According to an embodiment of the disclosure, an object recommendation method, medium, device and computing equipment are provided.

In this context, it is to be understood that the terminology which has been used is intended to be in the nature of words of description rather than of limitation. Moreover, any number of elements in the drawings are by way of example and not by way of limitation, and any nomenclature is used solely for differentiation and not by way of limitation.

The data referred to in the present disclosure may be data authorized by a user or sufficiently authorized by each party, and the embodiments/examples of the present disclosure may be combined with each other.

The following are descriptions of terms involved in this disclosure:

and (3) searching request: the search system usually refers to the text input by the user, and can also be a combination of the text input by the user and the label text automatically added by the server.

Pushing an object: in the scheme, at least audio, video, comment or user content fed back by the server based on the search request is mainly used.

The principles and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure.

Summary of The Invention

The inventor finds that, in the prior art, when a user needs to search a specific object, generally, after a search word is input in a search bar of a search system, a server returns an object list based on the search word, the user clicks any one of push objects in the object list, the content of the object is browsed by entering a detail page of the push object, and when the user wants to browse other content, the content of a next object can be browsed through a sliding operation or other switching operations. However, the next object presented to the user by the server is usually an object behind the push object operated by the user interactively in the object list obtained based on the search term, that is, the next object is also one of the object lists returned by the server based on the search term, and therefore, the next object cannot usually reflect the interest tendency expressed by the user by clicking the push object, which results in insufficient user experience.

According to the scheme, the second search request is generated through the label of the first object based on interactive operation and the first search request input in the search bar, and then the target object used for pushing is re-determined based on the second search request and serves as the next object browsed by the user, so that the next object can effectively reflect the interest tendency expressed by the user through interactive operation, and the user experience is improved.

Having described the general principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

Application scene overview

Referring to fig. 1, in content recommendation, a user sends a first search request 101 to a server 110 through a client 100, the server 110 feeds back at least one first object 102 to the client 100 based on the first search request 101, when the user enters a detail page of the first object 102 after performing an interaction operation on the first object 102 through the client 100, the server 110 generates a second search request 103 based on a tag of the first object 102 performing the interaction operation and the first search request 101, and the server 110 returns a target object 104 based on the second search request 103 as a next object browsed by the user to complete an object recommendation process.

It should be noted that, in the scenario shown in fig. 1, the client, the server, the first object, the first search request, the second search request, and the target object are only illustrated as an example, but the disclosure is not limited thereto, that is, the number of the client, the server, the first object, the first search request, the second search request, and the target object may be any.

Exemplary method

A method for object recommendation according to an exemplary embodiment of the present disclosure is described below with reference to fig. 2 to 5 in conjunction with the application scenario of fig. 1. It should be noted that the above application scenarios are merely illustrated for the convenience of understanding the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. Rather, embodiments of the present disclosure may be applied to any scenario where applicable.

Fig. 2 is a flowchart of an object recommendation method according to an embodiment of the present disclosure. As shown in fig. 2, the object recommendation method provided in this embodiment includes the following steps:

step S201, in response to the received first search request, acquiring a first object for pushing.

Wherein at least one of the first objects is present.

Specifically, the search object corresponding to the first search request may be of different categories, such as video, audio, comment, and user.

The first object returned based on the first search request may also have different specific categories corresponding to the different categories. Such as corresponding to video, the first object may be a single video, a collection of videos, a video uploader, etc., and corresponding to audio, the first object may be audio, video (e.g., an MV corresponding to audio), a song sheet, a singer, an album, etc.; corresponding to the comment, the first object can be a single comment, and also can be a comment topic, a comment tag, a comment article, a comment author and the like; the first object may be a single user, or may be a group of users, a discussion group, etc. corresponding to the user.

The content of the first search request may be keywords/words (e.g., "pop songs") or may be descriptive fields (e.g., "fun scenes seen at sporting events").

The server retrieves the corresponding at least one first object according to the first search request sent by the client, and returns the first object to the client in a list form.

Step S202, responding to the interactive operation of the first object, and generating a second search request.

And generating the second search request according to the first search request and the label of the first object corresponding to the interactive operation.

Specifically, each object stored in the server includes a corresponding tag (for example, audio may have a corresponding singer, title, and the like as tags, video may have a tag added manually or generated automatically, a keyword of a text of the comment content is used as a tag, and a text of a user such as an ID or a user name is used as a tag), and in the server, the tag is stored as an attribute of the object to reflect information such as the content and the feature of the first object.

Thus, when the server returns the first objects, the tags of the respective first objects are determined simultaneously. When the user performs an interactive operation (such as a click operation) on any first object, the server synchronously determines the label of the first object.

The server may obtain the second search request by combining the text in the tag of the first object with the text in the first search request, and if the first search request is "classic old song", and the tag of the first object (assumed to be a song of 90 years) is "90 years", "male singer", "old", then the second search request may be "classic old song/90 years/male singer/old".

Further, the second search request is automatically generated and sent to the server when the client sends the interactive operation to the server or when the client performs the sliding operation again (or another preset operation of switching the current display object) after the interactive operation, and the target object corresponding to the second search request is returned to the client by the server and displayed when the client receives the sliding operation (or another preset operation of switching the current display object).

Step S203, acquiring a second object to be pushed based on the second search request.

Wherein at least one of the second objects is present.

Specifically, the second object is a plurality of objects that the server retrieves from the objects of the inventory again based on the second search request.

The second object is only the object to be pushed determined by the server side, and is not directly pushed to the client, and the target object finally pushed to the client can be obtained only after screening.

And S204, determining a target object for pushing in the second object based on the second search request characteristic, the second object characteristic and the user characteristic and the object scoring model corresponding to the second search request.

Specifically, when the server receives the second search request and searches for the second object, the server may automatically extract feature information of the second search request and feature information of the second object according to the second search request and the second object, obtain user features of a user corresponding to the second search request, and then input the feature information of the second search request, the feature information of the second object, and the user features into a pre-trained object scoring model, where the object scoring model may input a score corresponding to the second object based on the input content, and may screen out one or more target objects pushed to the user according to the ranking of the scores.

Further, the training method of the object scoring model may be the same as or different from the training method of the object scoring model shown in the embodiment of fig. 7 and/or fig. 8, and is not limited herein.

Further, the feature information is information reflecting features of the second object, a probability of the second search request being invoked, a user usage or access frequency, and the like, and the user feature is information reflecting a user search and interactive operation tendency.

The feature information is a parameter extracted/calculated in advance by the server, and the corresponding feature information and the user feature can be obtained as long as the second search request, the second object and the corresponding user are determined.

Corresponding to each second object in the stock, the corresponding characteristic information can be obtained as long as the second object is determined; corresponding to each second search request, the server records the called times and the push result returned after calling, so that the server can also determine the corresponding characteristic information as long as the server receives the second search request; the server also stores corresponding operation records in real time corresponding to each user characteristic, and when receiving the second search request, the server can also directly determine the user characteristic corresponding to the second search request.

Further, after determining the target object, the server returns the target object to the client for the user to use.

According to the object recommendation method, a first object used for pushing is obtained by responding to a received first search request; generating a second search request in response to the interactive operation on the first object; acquiring a second object to be pushed based on the second search request; and determining a target object for pushing in the second object based on the second search request characteristic, the second object characteristic and the user characteristic and the object scoring model corresponding to the second search request. Therefore, the pushed target object can be determined again according to the first search request and the first object subjected to interactive operation, so that the target object can reflect the interest tendency expressed by interactive operation and the retrieval tendency expressed by the first search request at the same time, the requirements of users can be better met, and the user experience is improved.

Fig. 3 is a flowchart of an object recommendation method according to an embodiment of the present disclosure. As shown in fig. 3, the object recommendation method provided in this embodiment includes the following steps:

step S301, in response to the received first search request, acquiring a first object for pushing.

Wherein at least one of the first objects is present.

Specifically, the content of this step is the same as that of step S201 in the embodiment shown in fig. 2, and is not described here again.

Step S302, combining the text information in the first search request with the label of the first object to obtain an expanded statement.

In particular, the first search request and the second search request are in the form of sentences, and therefore, the sentence of the second search request needs to be redetermined in combination with the first search request and the tag of the first object,

after the first search request is combined with the tag of the first object, the search terms are increased, the search range is reduced, and if the search is directly performed, the situation that the required content cannot be searched easily occurs. In order to expand the search range, the expanded sentences obtained by combining the first search request and the tags of the first object need to be processed to obtain a plurality of expanded sentences, so that the expanded sentences can be searched together, and the purpose of simultaneously meeting the first search request and the interest tendency embodied by the interactive operation between the user and the first object can be effectively ensured.

Step S303, rewriting the expanded sentence based on the near-synonym and related word database to obtain a rewritten sentence containing the near-synonym and/or related word, and generating a second search request according to the rewritten sentence.

Specifically, the processing of the expanded sentence is mainly to obtain a plurality of rewritten sentences including the similar meaning words/related words based on the keyword in the expanded sentence based on the existing similar meaning words/related words model, and then to use the plurality of rewritten sentences together as the second search request, so that the server can perform the search through each of the plurality of rewritten sentences included in the second search request, and to use the search results obtained by all the rewritten sentences together as the second object.

And step S304, acquiring a second object to be pushed based on the second search request.

Wherein at least one of the second objects is present.

Specifically, the content of this step is the same as that of step S203 in the embodiment shown in fig. 2, and is not described here again.

Step S305, inputting the characteristics of the second search request, the characteristics of the at least one second object and the characteristics of the user corresponding to the second search request into an object scoring model to obtain the score of the at least one second object.

Specifically, corresponding to each rewritten sentence in the second search request, one corresponding feature information is obtained, and the combination of all feature information is the feature information of the second search request.

Further, the characteristics of the second search request comprise an identification code of the second search request, an object type to which the second search request is directed, and an interaction record of each second object obtained based on the second search request; the characteristics of the at least one second object comprise an identification code of the second object, a label of the second object and an interactive operation record of the second object in at least two periods; the characteristics of the user corresponding to the second search request comprise the interactive operation behavior preference characteristics of the user, interactive operation records of the user in at least two periods and historical search requests of the user.

Specifically, the server stores the corresponding identification code corresponding to each second search request (or each rewritten statement in the second search request), and directly calls the identification code when the second search request already exists; the type of object for which the second search request is directed can reflect the intent that the second search request needs to retrieve, such as for audio, video, or for singers, albums, songs; the interaction operation record of each second object obtained based on the second search request can reflect the number of interaction operations of the user corresponding to each second object in the second objects obtained based on the second search request in historical statistics, and further determine which second objects are more likely to be interested by the user.

The interactive operation records of the second object in at least two periods are used for representing the interactive operation statistical information of the second object in a long time (such as from the time of saving in the server) and the interactive operation statistical information in a real time (such as in the last 1 month), and the attention degree of the second object in the long time and the real time can be reflected through the two interactive operation records.

The user's interactive operation behavior preference feature is used for representing the sequence of objects which are interactively operated (such as clicked or commented) by the user, so that the long-time interest feature of the user can be reflected; the interactive operation records of the user in at least two periods are used for representing the interactive operation statistical information of the user in a long time (such as 1 year or since the account is registered) and the interactive operation statistical information in real time (such as in the last 1 month), so that the liveness condition of the user can be reflected.

In some embodiments, the features input into the object scoring model further include environmental features, and in this case, the second object score is obtained by:

and inputting the characteristics of the second search request, the characteristics of at least one second object, the environmental characteristics corresponding to the second search request and the characteristics of the user corresponding to the second search request into an object scoring model to obtain the score of the at least one second object.

Wherein the environmental characteristics include preference characteristics of the user for the object of the interactive operation on the set date and the set period.

In particular, the environmental characteristics are used to represent consumption tendency differences of each object in different environments/scenes. The preference characteristics of the user on the interactive operation object on the set date are used for reflecting the statistical difference of the user on the interactive operation on the same object on a working day/holiday (or different holidays); the preference characteristics of the user to the interactive operation object in the set time period are used for reflecting the statistical difference of the interactive operation of the user to the same object in the work/leisure time period or the early/late time period.

Further, as shown in fig. 4a, it is a schematic structural diagram of the object scoring model. The object scoring model 400 comprises a cross layer 410, a first full connection layer 420, a full splicing layer 430, a second full connection layer 440 and an output layer 450, wherein output ends of the cross layer 410 and the first full connection layer 420 are respectively connected with input ends of the full splicing layer 430, output ends of the full splicing layer 430 are connected with the output layer 450 sequentially through the second full connection layer 440, input ends of the cross layer 410 and the first full connection layer 420 are used for inputting data, the output layer 450 is used for outputting final results, the number of the first full connection layer 420 is at least two, and the number of the second full connection layer 440 is at least two.

Specifically, the cross layer 410 is configured to receive input discrete data and perform cross processing, the first full connection layer 420 receives input continuous data, and the full splicing layer 430 is configured to splice data output by the cross layer 410 and the first full connection layer 420; the second full-connect layer 440 connects the full-splice layer 430 and the output layer 450; the output layer 450 is used to output an output result based on the input discrete data and continuous data.

As shown in fig. 4b, which is a flow chart for calculating a second object score by the object scoring model. With reference to fig. 4a and 4b, the process of obtaining the second object score includes the following steps:

step S3051, performing first preprocessing on discrete data in the characteristics of the second search request, the characteristics of at least one second object and the characteristics of the user corresponding to the second search request, and performing second preprocessing on continuous data.

The first preprocessing comprises coding processing and sampling superposition processing, and the second preprocessing comprises splicing processing.

Specifically, the discrete data, such as the identification code of the second search request, the identification code of the second object, time, and the interactive operation behavior preference feature of the user, are encoded first, and then the encoded discrete data are subjected to sampling and superimposing processing, that is, processed by the existing Pooling algorithm.

Continuous data such as interactive operation records of the user in at least two periods, interactive operation records of the second object in at least two periods, and the like, and the data can be directly subjected to splicing processing.

And S3052, inputting the discrete data subjected to the first preprocessing into a cross layer for cross processing.

Specifically, the intersection processing may be cartesian intersection processing, or may be other intersection processing methods (such as inner product and outer product). Discrete data are integrated by performing cross processing on the discrete data, and are spliced with continuous data.

And S3053, inputting the continuous data subjected to the second preprocessing into the first full connection layer.

Specifically, the continuous data after the splicing processing is input into the first full connection layer for further processing. The number of the first full connection layers can be multiple, and continuous data can sequentially pass through each first full connection layer to be processed for multiple times.

And S3054, inputting the output result of the cross layer and the output result of the first full-connection layer into the full-splicing layer for splicing.

Specifically, the cross-processed discrete data and the processed continuous data output by the first full-link layer are subjected to full-splicing processing in the full-splicing layer and then input into the second full-link layer.

And S3055, enabling the result after splicing processing of the full splicing layer to pass through the second full connection layer and the output layer, and obtaining an output result through the output layer.

Specifically, the data after the full splicing processing is processed sequentially through the plurality of second full connection layers, and finally, an output result is obtained through an output function of the output layer.

And S3056, taking the output result as the score of the at least one second object.

Specifically, the end of the output layer may be a softmax function of a binary class, which is used to map the calculation result of the output layer into an interval of 0 to 1 as the output result. Since each second object can evaluate the probability that the second object meets the interest tendency of the user under the condition of the second search request and the corresponding user by a numerical value between 0 and 1, the output result of the output layer can be used as the score of the second object.

Step S306, determining a target object for pushing based on the score of the at least one second object.

Specifically, when only one target object is needed, a second object with the highest score output by the object scoring model may be directly used as the target object. When a plurality of target objects are needed, the target objects used for pushing are determined according to the ranking of the scores output by the object score model.

According to the object recommendation method of the embodiment of the disclosure, a first object for pushing is acquired by responding to a received first search request; combining text information in the first search request with a label of the first object to obtain an expanded statement, rewriting the expanded statement to obtain a plurality of rewritten statements which are jointly used as a second search request to obtain a second object to be pushed; inputting the characteristics of the second search request, the characteristics of the second object and the user characteristics corresponding to the second search request into an object scoring model, and determining a target object for pushing based on the output score. Therefore, a plurality of second objects can be obtained, and the target object which can reflect the interest tendency expressed by the interactive operation and the retrieval tendency expressed by the first search request can be determined according to the characteristics of the user and the characteristics of the second search request, so that the requirements of the user can be met better, and the user experience is improved.

Fig. 5 is a flowchart of an object recommendation method according to an embodiment of the present disclosure. As shown in fig. 5, the object recommendation method provided in this embodiment includes the following steps:

step S401, in response to the received first search request, acquiring a first object for pushing.

Wherein at least one of the first objects is present.

This step is the same as step S201 in the embodiment shown in fig. 2, and is not described here again.

Step S402, determining the label of the first object by querying a label database established in advance.

Specifically, the server may synchronously return the tag of the first object while acquiring the first object, and the tag of the first object is determined by querying a pre-established tag database.

In some embodiments, as shown in fig. 6, it is a flow chart of the establishment of the tag database. The tag database is determined by:

step S4021, respectively acquiring data to be extracted, which are stored by the stock object according to different contained modes.

The modalities comprise pictures, videos, audios and characters.

Specifically, since the objects in the stock usually contain data of multiple modalities, the modality of the data to be extracted needs to be determined first. For example, the video data further comprises text description and label content, and the audio data further comprises picture cover content and the like.

S4022, extracting data to be extracted in corresponding modes through a convolutional neural network to obtain features to be extracted in different modes.

In particular, convolutional neural networks are existing models for extracting data of different types of modalities.

Specifically, obtaining the data to be extracted in different modalities includes at least one of the following:

case one (not shown), scene and face feature vectors in video frames obtained by extracting stock video samples based on a two-dimensional convolutional neural network are taken as feature vectors of D1 dimension of the stock video.

Specifically, for each inventory object, video frames are sampled according to a set rule (for example, one frame is extracted every 1 second), then a two-dimensional convolutional neural network is used for respectively extracting semantic information of scenes and faces of each frame, finally, the output of the penultimate layer of the full-link layer of the convolutional neural network is used as feature expression, NxD 1-dimensional feature vectors are obtained, N represents N frames, and D1 represents that the feature dimension of each frame is D1 dimensions. And finally, overlapping NxD 1-dimensional feature vectors of all frames to obtain a D1-dimensional feature vector which is used as the feature of the video, and further obtaining a D1-dimensional scene feature and a D1-dimensional face feature vector.

Further, the two-dimensional convolutional neural network may select a residual error network VGG to extract scene and face feature vectors, or may select any other two-dimensional convolutional neural network.

Case two (not shown), video features of the inventory video are extracted based on the three-dimensional convolutional neural network as feature vectors of D2 dimension of the inventory video.

Specifically, the video in stock is input into a three-dimensional convolutional neural network, video semantic information is extracted, the output result of the penultimate layer of the fully-connected layer of the three-dimensional convolutional neural network is used as a feature vector of the whole video, and the feature dimension of the feature vector is recorded as D2 dimension.

Further, the three-dimensional convolutional neural network may be a 3DCNN network, or any other three-dimensional convolutional neural network.

Case three (not shown), the language processing based convolutional neural network extracts the text features of the stock video as the feature vector of D3 dimension of the stock video.

Specifically, the video or the picture is input into a language processing convolution network, text features are extracted, the output result of the language processing convolution network is used as a feature vector of the text, and the corresponding feature dimension is recorded as D3 dimension.

Furthermore, the language processing convolutional neural network can be a Bert network, and can also be any other convolutional neural network used for language extraction in videos/pictures.

Case four (not shown), the audio features of the inventory video/audio are extracted based on the audio processing convolutional neural network as the feature vector of D4 dimension of the inventory video.

Specifically, the video or audio is input into an audio processing convolutional neural network, audio features are extracted, and the corresponding feature dimension is recorded as D4 dimension.

Further, the audio processing convolutional neural network may be a vggish network, but is not limited to the convolutional neural network in practical application.

And S4023, splicing the features to be extracted to obtain a feature vector corresponding to the stock object.

And splicing to obtain the feature vector of the D-dimensional inventory video according to the following rules: D2D 1+ D2+ D3.

By directly splicing the dimensional features, a feature vector reflecting the overall features of the object can be obtained.

Step S4024, determining corresponding classifications of the feature vectors based on the pre-trained object classifier.

Specifically, the object classifier is configured to determine a category to which the inventory object conforms based on the feature vector, and since the same inventory object may belong to multiple categories (for example, a video may simultaneously conform to a "landscape" category, a "high definition" category, a "beach" category, etc.), the category to which each inventory object conforms may be used as a tag of the inventory object.

Further, the pre-trained object classifier is obtained by the following method:

step one (not shown), the kind of label contained in the object is determined based on the service scene.

Specifically, the tag categories are usually designed and divided by the administrator according to the needs of the business scenario. Labels such as music, including singers, years, and songs; the video tags include life categories, popular science categories, and the like.

And step two (not shown), taking the object added with the label category and the corresponding feature vector as a training set and a verification set.

Specifically, a part of objects used for training is added with corresponding labels in advance, and then the object and one part of the calculated feature vectors of the object are used as a training set, and the other part is used as a verification set, so as to train the object classifier.

And step three (not shown), inputting the training set and the verification set into an object classifier, and obtaining the pre-trained object classifier through deep learning training.

Specifically, the method for training the object classifier is the same as the training method of the conventional neural network, and the object classifier capable of identifying the object type can be obtained through deep learning.

In some embodiments, there may be multiple object classifiers, each object classifier identifying an object class of a modality (e.g., an object classifier that specifically identifies a user and an object classifier that specifically identifies audio), and combining the identification results of the multiple object classifiers to collectively serve as an object classifier for identifying an inventory object class.

Further, after determining the corresponding classification of the feature vectors, the method further includes:

and determining that the association degree of the inventory object and the corresponding classification meets the set conditions based on the set auditing rule.

Specifically, the degree of association between the object type calculated by the object classifier and the inventory object may be insufficient, for example, the association between the football-type video and the "travel" classification is low, and at this time, it may be determined that the degree of association between the inventory object and the corresponding classification satisfies the set condition through the set auditing rule (e.g., manual auditing, or existing association algorithm).

Step S4025, all the classifications contained in the stock object are used as the labels of the stock object and stored in the label database.

Specifically, since the inventory object usually belongs to a plurality of categories, each of the categories may be used as a label of the inventory object, and the labels of the inventory object may be obtained by combining all the labels. And then the label is saved in a label database so as to be convenient for inquiring and calling at any time.

Step S403, generating a second search request in response to the interactive operation on the first object.

And generating a second search request according to the first search request and the label of the first object corresponding to the interactive operation.

And step S404, acquiring a second object to be pushed based on the second search request.

Wherein at least one of the second objects is present.

Specifically, steps S403 to S404 are the same as steps S202 to S203 in the embodiment shown in fig. 2, and are not repeated here.

Step S405, based on a semantic correlation algorithm, filtering out at least one second object, wherein the correlation between a first object corresponding to the interactive operation and the first search request is lower than a set value.

Specifically, the semantic relevance algorithm can calculate the relevance between the first object corresponding to the interactive operation and the text of the first search request by identifying the label or feature information of the second object, and if the relevance is too low, even if the content is very hot, the problem that the interest tendency of the user is not met may exist.

Step S406, determining a target object for pushing in the second object based on the second search request characteristic, the second object characteristic and the user characteristic and the object scoring model corresponding to the second search request

Specifically, the content of step S406 is the same as that of step S204 in the embodiment shown in fig. 2, and is not repeated here.

According to the object recommendation method, a first object used for pushing is obtained by responding to a received first search request; obtaining the label combination of a first object by inquiring a label database trained in advance, and obtaining a second search request by combining the label of the first object with the first search request so as to obtain a second object to be pushed; inputting the characteristics of the second search request, the characteristics of the second object and the user characteristics corresponding to the second search request into an object scoring model, and determining a target object for pushing based on the output score. Therefore, the first object and the target object used for pushing can be determined according to the label of the inventory object, so that the pushed content can better meet the requirements of users, and the user experience is improved.

Fig. 7 is a flowchart of a training method for an object scoring model according to an embodiment of the present disclosure. As shown in fig. 7, the object scoring model training method provided in this embodiment includes the following steps:

step S501, a data set to be extracted and a labeled data set are obtained based on a client log and a server log in a server.

The category of the data set to be extracted comprises a second search request feature, a user feature and a second object feature, the labeled data set comprises labeled data of user interaction operation, the second search request is generated according to the first search request and a label of a first object corresponding to the interaction operation, the first object is obtained according to the first search request, and the second object is obtained according to the second search request.

Specifically, the second search request feature, the user feature, the object feature, and the interactive operation record are data stored in a client log and a server log of the server, where the second search request feature and the object feature are information reflecting features such as a probability of being called, a user use frequency, or an access frequency of the second search request/object, the user feature is information reflecting a user search and an interactive operation tendency, and the labeled data of the interactive operation is information reflecting an interactive operation tendency of the user on the object (for example, when the user performs a viewing/listening trial on a push object for more than a set time period, performs a positive interactive operation such as a click, a comment, a forward, a share, or the like, the labeled data of the interactive operation may be a tag value of 1, and otherwise, the labeled data of the interactive operation is 0).

The second search request is a new search request generated on the basis of the interactive operation of the first search request and the first object obtained by the user from the first search request.

And collecting data corresponding to all the second search request characteristics, the user characteristics, the object characteristics and the interactive operation records, namely the data set to be extracted.

The second object corresponding to the second object feature may be the first object (or at least one of the plurality of first objects), or may not be the first object (in this case, the second object is at least one object recalled from objects in the server inventory and obtained based on the second search request recorded in the client log and the server log).

And S502, extracting corresponding data characteristics according to the category of the data set to be extracted.

Specifically, the data set to be extracted includes different types of feature data and interactive operation records, so that the data features of the corresponding data need to be extracted in different extraction manners, so as to ensure that the extracted data features can accurately and effectively reflect the corresponding features.

And S503, preprocessing the data characteristics, and inputting the preprocessed data characteristics into an object scoring model to obtain a prediction score.

Specifically, since the data in the data set to be extracted includes different types and respectively corresponds to the second search request, the inventory object, the user, and the interactive operation, even after the data is extracted, the data cannot be directly used for training, and therefore, the extracted data features need to be preprocessed on the basis of corresponding to each object and then input into the object scoring model. The interaction operation corresponding to the labeled data set is compared with the output prediction score of the object scoring model so as to evaluate the data characteristics of the training result of the object scoring model.

And step S504, calculating a loss function based on the prediction score and the labeled data, and training an object score model based on the loss function.

Specifically, the accuracy of the output result of the object scoring model can be ensured by comparing the prediction scoring with the actual interactive operation data (or the labeled data) obtained by labeling the data set, taking the comparison result as a loss function, and training the object scoring model by iterative training of the loss function.

According to the object scoring model training method, a data set to be extracted and a labeled data set are obtained based on a client log and a server log in a server, corresponding data features are extracted according to the category of the data set to be extracted, the data features are preprocessed and input into an object scoring model to obtain a prediction score, and the object scoring model is trained based on a loss function. Therefore, the object scoring model can evaluate the interest degree of the object correspondingly input into the object scoring model by the user based on the second search request feature data, the object feature data and the user feature data in the client log and the server log, and can optimize the evaluation accuracy based on the interactive operation record data, so that the accuracy of the output result of the object scoring model is effectively ensured when the content is recommended to the user, the pushed content can be better satisfied with the requirements of the user, and the user experience is improved.

Fig. 8a is a schematic structural diagram of an object scoring model according to an embodiment of the present disclosure. As shown in fig. 8a, the object scoring model 800 includes a cross layer 810, a full connection layer 820 and an output layer 830, wherein the outputs of the cross layer 810 and the full connection layer 820 are respectively connected to the inputs of the output layer 830, the inputs of the cross layer 810 and the full connection layer 820 are used for inputting data, the output layer 830 is used for outputting the final result, and at least two full connection layers 820 are provided.

Specifically, the output layer 830 includes a full-concatenation layer to perform full-concatenation processing on data output by the cross layer 810 and the full-concatenation layer 820.

Fig. 8b is a flowchart of a training method of an object scoring model according to an embodiment of the present disclosure. As shown in fig. 8a and 8b, the method for training the object score model according to this embodiment includes the following steps:

step S601, obtaining a data set to be extracted and a labeled data set based on a client log and a server log in a server.

Specifically, the content of this step is the same as that of step S501 in the embodiment shown in fig. 7, and details are not repeated here.

In some embodiments, in the process of obtaining the data set to be extracted based on the client log and the server log in the server, the category of the data set to be extracted includes the second search request feature, the user feature, the object feature, the environmental feature, and the user interaction record, that is, the category of the data set to be extracted also includes the environmental feature.

In particular, the environmental characteristics are used to represent consumption tendency differences of each object in different environments/scenes.

Step S602, according to the category of the data set to be extracted, extracting the granularity characteristic of category data.

Specifically, the granular features are used to indicate a specific feature type of each category of data, for example, the granular features corresponding to the second search request feature include an identification code of the second search request, an object type to which the second search request is directed, and an interaction record of each second object obtained based on the second search request, and the granular features of other categories of data refer to step S305 in the embodiment shown in fig. 3.

By extracting the granularity characteristic of each category of data and extracting/calculating the specific data of each granularity characteristic, the data characteristic in the data set to be extracted can be obtained.

And step S603, adding corresponding zone bit identification codes for the data and the granularity characteristics in the data set to be extracted.

Specifically, corresponding to each calculated/extracted granularity feature and corresponding data, the server adds a corresponding zone bit identification code and stores the zone bit identification code in the server so as to perform object scoring model training and quickly call in practical application.

And step S604, performing first preprocessing on discrete data in the data characteristics.

Wherein the first preprocessing comprises coding processing and sample superposition processing.

Specifically, the user feature, the second search request feature, and the object feature include both discrete data and continuous data, and the environment feature is discrete data. By processing discrete data and continuous data respectively, various data can be sampled and spliced into object-based training data.

And step S605, performing second preprocessing on the continuous data in the data characteristics.

Wherein the second pre-processing comprises a splicing process.

Specifically, the continuous data is directly input to the full connection layer for splicing. And step S606, inputting the discrete data after the first preprocessing into an intersection layer for intersection processing.

Specifically, the discrete data after the encoding processing and the sampling superposition processing may be sent to a cross layer for cross processing, where the cross processing may be cartesian cross processing, or may also be a cross processing manner such as an inner product, an outer product, or the like.

And step S607, inputting the continuous data after the second preprocessing into the full connection layer.

Specifically, the continuous data will pass through at least two fully connected layers and be processed sequentially.

And step S608, inputting the output result of the cross layer and the output result of the full connection layer into the output layer, and obtaining the output result through the output layer.

Specifically, the discrete data after the cross processing and the continuous data after the splicing processing can be input to an output layer for full splicing processing and other further processing, so as to obtain an output result.

And step S609, taking the output result as the prediction score of the second object.

Specifically, the training sample is all second search requests corresponding to a single push object, users corresponding to each second search request, and a record of whether the users perform interactive operations on the push object. By inputting these training samples to the cross layer, the full link layer, and the output layer in order after preprocessing, the prediction score p corresponding to each training sample can be obtained.

And step S610, calculating a loss function L based on the prediction scores and the labeling data.

Specifically, the calculation formula is as follows: l ═ ylog (p) + (1-y) log (1-p), where y is the interaction data in the user interaction data set corresponding to the training sample.

Corresponding to each training sample, a corresponding y value of the data sample input to the output layer of the deep neural network can be calculated, a corresponding prediction score p can be calculated, and the value of the corresponding loss function L can be calculated by substituting y and p into a formula.

Step S611, optimizing the object scoring model based on the iterative training of the loss function L to obtain a pre-trained object scoring model.

Specifically, by performing iterative training on the loss function L (for example, performing iterative training with a target smaller than a set value), each parameter in the object scoring model is optimized, so that the object scoring model can accurately calculate the interest degree (or the probability of performing an interactive operation) of the user on the pushed object based on the input second search request.

According to the object scoring model training method, a data set to be extracted and labeled data are obtained based on a client log and a service end log in a server, then the granularity characteristics of category data are extracted according to the category of the data set to be extracted, corresponding mark bit identification codes are added for the data and the granularity characteristics in the data set to be extracted, discrete data and the discrete data are preprocessed and input into an object scoring model to obtain a prediction score, and the object scoring model is trained based on a loss function. Therefore, the object scoring model can accurately evaluate the probability of interactive operation of the object input into the object scoring model by the user based on the second search request feature data, the object feature data and the user feature data, so that the accuracy of the output result of the object scoring model is effectively ensured when the content is recommended to the user, the requirement of the user can be better met by the pushed content, and the user experience is improved.

Exemplary Medium

Having described the method of the exemplary embodiment of the present disclosure, next, a storage medium of the exemplary embodiment of the present disclosure will be described with reference to fig. 7.

Referring to fig. 9, a program product 90 for implementing the above method according to an embodiment of the present disclosure is described, which may employ a portable compact disc read only memory (CD-ROM) and include program code, and may be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto.

The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

A readable signal medium may include a propagated data signal with readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. The readable signal medium may also be any readable medium other than a readable storage medium.

Program code for carrying out operations of the present disclosure may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user computing device, partly on the user device, partly on a remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN).

Exemplary devices

After introducing the media of the exemplary embodiment of the present disclosure, next, an object recommendation apparatus of the exemplary embodiment of the present disclosure is described with reference to fig. 8, which is used for implementing the object recommendation method in any of the method embodiments described above, and the implementation principle and the technical effect are similar, and are not described herein again.

The present disclosure provides an object recommendation device 1000, including:

a first obtaining module 1010, configured to obtain, in response to a received first search request, first objects for pushing, where at least one of the first objects is obtained;

a generating module 1020, configured to generate a second search request in response to an interactive operation on the first object, where the second search request is generated according to the first search request and a tag of the first object corresponding to the interactive operation;

a second obtaining module 1030, configured to obtain, based on the second search request, at least one second object to be pushed;

the determining module 1040 is configured to determine, based on the second search request feature, the second object feature, and the user feature and the object scoring model corresponding to the second search request, a target object for pushing in the second object.

Optionally, the determining module 1040 is specifically configured to: combining text information in the first search request with a tag of the first object to obtain an expanded statement; and rewriting the expanded sentence based on the near-synonym and related word database to obtain a rewritten sentence containing the near-synonym and/or related words, and generating a second search request according to the rewritten sentence.

Optionally, the determining module 1040 is specifically configured to: inputting the characteristics of the second search request, the characteristics of at least one second object and the characteristics of the user corresponding to the second search request into an object scoring model to obtain the score of the at least one second object; determining a target object for pushing based on the score of the at least one second object.

Optionally, the determining module 1040 is specifically configured to, when the object scoring model includes a cross layer, a first full connection layer, a full splicing layer, a second full connection layer, and an output layer, output ends of the cross layer and the first full connection layer are respectively connected to an input end of the full splicing layer, an output end of the full splicing layer is connected to the output layer sequentially through the second full connection layer, input ends of the cross layer and the first full connection layer are used to input data, the output layer is used to output a final result, there are at least two first full connection layers, and there are at least two second full connection layers; inputting the second search request characteristic, the at least one second object characteristic and the user characteristic corresponding to the second search request into an object scoring model to obtain a score of the at least one second object, wherein the score of the at least one second object comprises the following steps: performing first preprocessing on discrete data in the second search request characteristic, at least one second object characteristic and the user characteristic corresponding to the second search request, and performing second preprocessing on continuous data, wherein the first preprocessing comprises coding processing and sampling superposition processing, and the second preprocessing comprises splicing processing; inputting the discrete data after the first pretreatment into a cross layer for cross treatment; inputting the continuous data after the second preprocessing into a first full connection layer; inputting the output result of the cross layer and the output result of the first full-connection layer into the full-splicing layer for splicing; the result after the splicing processing of the full splicing layer passes through the second full connecting layer and the output layer, and an output result is obtained through the output layer; and taking the output result as the score of the at least one second object.

Optionally, the determining module 1040 is further configured to: inputting the characteristics of the second search request, the characteristics of at least one second object, the environmental characteristics corresponding to the second search request and the characteristics of the user corresponding to the second search request into an object scoring model to obtain the score of the at least one second object; and determining a target object for pushing based on the score of the at least one second object.

Optionally, the determining module 1040 is specifically configured to: the environment characteristics include a preference characteristic of the user for the object of the interactive operation on a set date and a set period.

Optionally, the determining module 1040 is specifically configured to: the characteristics of the second search request comprise an identification code of the second search request, the type of the object to which the second search request aims, and an interactive operation record of each second object obtained based on the second search request; the characteristics of the at least one second object comprise an identification code of the second object, a label of the second object and an interactive operation record of the second object in at least two periods; the characteristics of the user corresponding to the second search request comprise the interactive operation behavior preference characteristics of the user, interactive operation records of the user in at least two periods and historical search requests of the user.

Optionally, the generating module 1020 is specifically configured to determine the tag of the first object by querying a pre-established tag database.

Optionally, the generating module 1020 is specifically configured to determine the tag database by: respectively acquiring data to be extracted, which are stored by the inventory object according to different contained modes, wherein the modes comprise pictures, videos, audios and characters; extracting data to be extracted in corresponding modes through a convolutional neural network to obtain features to be extracted in different modes; splicing the features to be extracted to obtain a feature vector corresponding to the stock object; determining a corresponding classification of the feature vectors based on a pre-trained object classifier; and all the classifications contained in the inventory object are used as the labels of the inventory object and are stored in the label database.

Optionally, the generating module 1020 is specifically configured to obtain the pre-trained object classifier by: determining the type of a label contained in the object based on the service scene; taking the video added with the label variety and the corresponding feature vector as a training set and a verification set; and inputting the training set and the verification set into an object classifier, and obtaining the pre-trained object classifier through deep learning training.

Optionally, the second obtaining module 1030 is further configured to, after obtaining the second object to be pushed based on the second search request, filter, based on a semantic relevance algorithm, a first object corresponding to the interactive operation in the at least one second object and a second object whose relevance of the first search request is lower than a set value.

Referring to fig. 11, an object scoring model training apparatus according to an exemplary embodiment of the present disclosure is described, for implementing an object scoring model training method in any of the method embodiments described above, which is similar to the method embodiments described above in terms of implementation principle and technical effect and is not described herein again.

The present disclosure provides an object scoring model training device 1100, comprising:

an obtaining module 1110, configured to obtain a to-be-extracted data set and a labeled data set based on a client log and a server log in a server, where a category of the to-be-extracted data set includes a second search request feature, a user feature, and a second object feature, the labeled data set includes labeled data of user interaction, the second search request is generated according to a first search request and a tag of a first object corresponding to the interaction, the first object is obtained according to the first search request, and the second object is obtained according to the second search request;

an extracting module 1120, configured to extract corresponding data features according to the category of the data set to be extracted;

a processing module 1130, configured to input the preprocessed data features into an object scoring model to obtain a prediction score;

a training module 1140 for calculating a loss function based on the prediction scores and the annotation data and training the object score model based on the loss function.

Optionally, the obtaining module 1110 is specifically configured to obtain the second search request by: combining text information in the first search request with a label of a first object corresponding to the interactive operation to obtain an expanded statement; and rewriting the expanded sentence based on the near-synonym and related word database to obtain a rewritten sentence containing the near-synonym and/or related words, and generating a second search request according to the rewritten sentence.

Optionally, the extracting module 1120 is specifically configured to extract a granularity feature of category data according to a category of the data set to be extracted; and adding corresponding zone bit identification codes for the data and the granularity characteristics in the data set to be extracted.

Optionally, the processing module 1130 is specifically configured to, when the object scoring model includes a cross layer, a full connection layer, and an output layer, outputs of the cross layer and the full connection layer are respectively connected to an input end of the output layer, the input ends of the cross layer and the full connection layer are used for inputting data, the output layer is used for outputting a final result, and there are at least two full connection layers; performing first preprocessing on discrete data in the data characteristics, wherein the first preprocessing comprises coding processing and sampling superposition processing; performing second preprocessing on continuous data in the data characteristics, wherein the second preprocessing comprises splicing processing; inputting the discrete data after the first pretreatment into a cross layer for cross treatment; inputting the continuous data after the second pretreatment into a full connection layer, and obtaining an output result through an output layer; and taking the output result as the prediction score of the second object.

Optionally, the training module 1140 is specifically configured to calculate a loss function L based on the prediction score and the annotation data: l ═ ylog (p) + (1-y) log (1-p), where p is the prediction score and y is the annotation data, used to indicate whether the user and the second object in the training sample are interactive; and optimizing the object scoring model based on the iterative training of the loss function L to obtain a pre-trained object scoring model.

Optionally, the obtaining module 1110 is specifically configured to further include the environmental characteristic in the category of the data set to be extracted.

Exemplary computing device

Having described the methods, media, and apparatus of the exemplary embodiments of the present disclosure, a computing device of the exemplary embodiments of the present disclosure is next described with reference to fig. 12.

The computing device 120 shown in fig. 12 is only one example and should not place any limitation on the scope of use and functionality of embodiments of the present disclosure.

As shown in fig. 12, computing device 120 is embodied in the form of a general purpose computing device. Components of computing device 120 may include, but are not limited to: the at least one processing unit 1201 and the at least one storage unit 1202 may be coupled together via a bus 1203 to the various system components including the processing unit 1201 and the storage unit 1202.

The bus 1203 includes a data bus, a control bus, and an address bus.

The storage unit 1202 may include readable media in the form of volatile memory, such as Random Access Memory (RAM)12021 and/or cache memory 12022, and may further include readable media in the form of non-volatile memory, such as Read Only Memory (ROM) 12023.

The storage unit 1202 may also include a program/utility 12025 having a set (at least one) of program modules 12024, such program modules 12024 including, but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which, or some combination thereof, may comprise an implementation of a network environment.

Computing device 120 may also communicate with one or more external devices 1204 (e.g., keyboard, pointing device, etc.). Such communication may occur via input/output (I/O) interfaces 1205. Also, computing device 120 may communicate with one or more networks (e.g., a Local Area Network (LAN), a Wide Area Network (WAN), and/or a public network, such as the internet) through network adapter 1206. As shown in FIG. 4, network adapter 1206 communicates with the other modules of computing device 120 over bus 1203. It should be understood that although not shown in the figures, other hardware and/or software modules may be used in conjunction with computing device 120, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, among others.

It should be noted that although several units/modules or sub-units/modules of the object recommendation device and the object scoring model training device are mentioned in the above detailed description, such division is merely exemplary and not mandatory. Indeed, the features and functions of two or more units/modules described above may be embodied in one unit/module according to embodiments of the present disclosure. Conversely, the features and functions of one unit/module described above may be further divided into embodiments by a plurality of units/modules.

Further, while the operations of the disclosed methods are depicted in the drawings in a particular order, this does not require or imply that these operations must be performed in this particular order, or that all of the illustrated operations must be performed, to achieve desirable results. Additionally or alternatively, certain steps may be omitted, multiple steps combined into one step execution, and/or one step broken down into multiple step executions.

While the spirit and principles of the present disclosure have been described with reference to several particular embodiments, it is to be understood that the present disclosure is not limited to the particular embodiments disclosed, nor is the division of aspects, which is for convenience only as the features in such aspects may not be combined to benefit. The disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. An object recommendation method comprising the steps of:

responding to a received first search request, and acquiring first objects for pushing, wherein at least one first object is acquired;

responding to the interactive operation of the first object, and generating a second search request according to the first search request and the label of the first object corresponding to the interactive operation;

and determining a target object for pushing in the second object based on the second search request characteristic, the second object characteristic and a user characteristic and an object scoring model corresponding to the second search request.

2. The method of claim 1, the generating a second search request in response to the interaction with the first object, comprising:

combining the text information in the first search request with the label of the first object to obtain an expanded sentence;

and rewriting the expanded sentence based on the near-meaning word and related word database to obtain a rewritten sentence containing the near-meaning word and/or related words, and generating a second search request according to the rewritten sentence.

3. The method of claim 1, wherein determining a target object in the second object for pushing based on the second search request feature, the second object feature, and a user feature and an object scoring model corresponding to the second search request comprises:

inputting the characteristics of the second search request, the characteristics of at least one second object and the characteristics of a user corresponding to the second search request into the object scoring model to obtain a score of the at least one second object;

and determining a target object for pushing in the second object based on the score of at least one second object.

4. The method according to claim 3, wherein the object scoring model comprises a cross layer, a first full connection layer, a full splicing layer, a second full connection layer and an output layer, wherein the output ends of the cross layer and the first full connection layer are respectively connected with the input end of the full splicing layer, the output end of the full splicing layer is connected with the output layer through the second full connection layer in sequence, the input ends of the cross layer and the first full connection layer are used for inputting data, the output layer is used for outputting a final result, the number of the first full connection layer is at least two, and the number of the second full connection layer is at least two;

inputting the second search request characteristic, at least one second object characteristic and a user characteristic corresponding to the second search request into the object scoring model to obtain a score of at least one second object, wherein the score comprises the following steps:

performing first preprocessing on discrete data in the second search request characteristic, at least one second object characteristic and a user characteristic corresponding to the second search request, and performing second preprocessing on continuous data, wherein the first preprocessing comprises coding processing and sampling superposition processing, and the second preprocessing comprises splicing processing;

inputting the discrete data after the first preprocessing into the cross layer for cross processing;

inputting the continuous data after the second preprocessing into the first full-connection layer;

inputting the output result of the cross layer and the output result of the first full-connection layer into the full-splicing layer for splicing;

the result after splicing processing of the full splicing layer passes through the second full connecting layer and the output layer, and an output result is obtained through the output layer;

and taking the output result as the score of the at least one second object.

5. The method of claim 1, wherein determining a target object for pushing in the second object based on the second search request feature, the second object feature, and a user feature and an object scoring model corresponding to the second search request further comprises:

inputting the characteristics of the second search request, the characteristics of at least one second object, the environmental characteristics corresponding to the second search request and the characteristics of the user corresponding to the second search request into the object scoring model to obtain the score of the at least one second object;

determining a target object for pushing based on the score of at least one of the second objects.

6. The method of claim 5, the environmental characteristics comprising preference characteristics of the user for an interoperating object on a set date and a set period of time.

7. The method of claim 1, wherein the characteristics of the second search request comprise an identification code of the second search request, a type of object for which the second search request is directed, and an interaction record of each second object obtained based on the second search request; the characteristics of the at least one second object comprise an identification code of the second object, a label of the second object and an interaction record of the second object in at least two periods; the characteristics of the user corresponding to the second search request comprise the interactive operation behavior preference characteristics of the user, interactive operation records of the user in at least two periods and historical search requests of the user.

8. A method of training a subject scoring model, comprising:

obtaining a data set to be extracted and a labeled data set based on a client log and a server log in a server, wherein the category of the data set to be extracted comprises a second search request characteristic, a user characteristic and a second object characteristic, the labeled data set comprises labeled data of user interaction operation, the second search request is generated according to a first search request and a label of a first object corresponding to the interaction operation, the first object is obtained by pushing according to the first search request, and the second object is obtained according to the second search request;

preprocessing the data characteristics, and inputting the data characteristics into an object scoring model to obtain a prediction score;

and calculating a loss function based on the prediction score and the labeling data, and training the object score model based on the loss function.

9. A computer-readable storage medium, comprising: the computer-readable storage medium has stored therein computer-executable instructions for implementing the method of any one of claims 1 to 8 when executed by a processor.

10. A computing device, comprising:

at least one processor;

and a memory communicatively coupled to the at least one processor;

wherein the memory stores instructions executable by the at least one processor to cause the computing device to perform the method of any of claims 1 to 8.