CN112287225B

CN112287225B - Object recommendation method and device

Info

Publication number: CN112287225B
Application number: CN202011183756.6A
Authority: CN
Inventors: 丁敏; 崔正文; 黄浩; 袁晓洁
Original assignee: Beijing QIYI Century Science and Technology Co Ltd
Current assignee: Beijing QIYI Century Science and Technology Co Ltd
Priority date: 2020-10-29
Filing date: 2020-10-29
Publication date: 2023-09-08
Anticipated expiration: 2040-10-29
Also published as: CN112287225A

Abstract

The embodiment of the invention provides an object recommendation method and device, and relates to the technical field of data processing, wherein the method comprises the following steps: obtaining user characteristics of a target user and object characteristics of an object to be recommended, wherein the user characteristics and the object characteristics are respectively used as the target user characteristics and the target object characteristics; inputting the target user characteristics and the target object characteristics into a first recommendation prediction model trained in advance, and predicting whether to recommend an object to be recommended to a target user; and if the output result of the first recommendation prediction model represents that the object to be recommended can be recommended to the target user, recommending the object to be recommended to the target user. By applying the scheme provided by the embodiment of the invention to recommend the object, the object can be quickly and accurately recommended to the user.

Description

Object recommendation method and device

Technical Field

The present invention relates to the field of data processing technologies, and in particular, to an object recommendation method and apparatus.

Background

Users may tend to view objects of interest to themselves when using the network platform, e.g., video, news, pictures, merchandise, etc., which may be different for different users. For example, taking the example where the above object is a video, user 1 tends to view a video of a sports class that is longer in duration, user 2 tends to view a video of an entertainment class that is shorter in duration, and so on. The network platform has a large number of objects aiming at different contents, and for each user, only a part of the objects are interested by the user, and the user needs to spend a large amount of time for searching the objects interested by the user in the large number of objects, so that the time for searching the objects interested by the user can be shortened by rapidly and accurately recommending the objects interested by the user to the user, and the use experience of the user is improved.

However, because the number of users and objects in the network platform is large, more data needs to be processed in the process of recommending objects to the users, more computing resources need to be consumed if the accurate object recommendation to each user is realized, and when a large number of users need to be simultaneously recommended, the object recommendation efficiency is lower and the speed is lower. In view of the above, it is desirable to provide a method capable of recommending objects to a user quickly and accurately.

Disclosure of Invention

The embodiment of the invention aims to provide an object recommending method, so that objects can be quickly and accurately recommended to users. The specific technical scheme is as follows:

in a first aspect of the embodiment of the present invention, there is first provided an object recommendation method, including:

obtaining user characteristics of a target user and object characteristics of an object to be recommended, wherein the user characteristics and the object characteristics are respectively used as the target user characteristics and the target object characteristics;

inputting the target user characteristics and the target object characteristics into a pre-trained first recommendation prediction model, predicting whether to recommend the object to be recommended to the target user or not, wherein the first recommendation prediction model is obtained by jointly training with a second recommendation prediction model, the first recommendation prediction model is used for carrying out characteristic fusion on the characteristics of the user and the characteristics of the object included in input data to obtain a first fusion result, predicting whether to recommend the object to be recommended to the target user or not based on the first fusion result, the second recommendation prediction model is used for extracting association characteristics representing association relation between the characteristics of the user and the characteristics of the object included in the input data, carrying out fusion on the extracted association characteristics to obtain a second fusion result, predicting whether to recommend the object to be recommended to the target user or not based on the second fusion result, and adjusting parameters of the first recommendation model based on difference between the first fusion result and the second fusion result output by the first recommendation prediction model and difference between the second recommendation result output by the second recommendation prediction model in the joint training process so that the first recommendation result approaches to the first recommendation result and the second recommendation prediction model;

And if the output result of the first recommendation prediction model indicates that the object to be recommended can be recommended to the target user, recommending the object to be recommended to the target user.

In a second aspect of the embodiment of the present invention, there is also provided a model training method, where the method includes:

obtaining sample user characteristics of a sample user and sample object characteristics of a sample object;

inputting the characteristics of the sample user and the characteristics of the sample object into a first initial model, carrying out characteristic fusion on the characteristics of the sample user and the characteristics of the sample object included in the input data, and predicting whether to recommend the sample object to the sample user based on a first sample fusion result to obtain a first sample prediction result;

inputting the sample user features and the sample object features into a second initial model, extracting sample association features between the features of the user and the features of the object included in the input data, fusing the extracted sample association features, predicting whether to recommend the sample object to the sample user based on a second fusion result, and obtaining a second sample prediction result;

and adjusting model parameters of a first initial model and a second initial model based on the difference between the first sample fusion result and the second sample fusion result and the difference between the first sample prediction result and the second sample prediction result until a preset training ending condition is met, so that the first sample fusion result approaches to the second sample fusion result, the first sample prediction result approaches to the second sample prediction result, determining the trained first initial model as a first recommended prediction model for predicting whether to recommend an object to a user, and determining the trained second initial model as a second recommended prediction model for predicting whether to recommend the object to the user.

In a third aspect of the embodiment of the present invention, there is also provided an object recommendation apparatus, where the apparatus includes:

the feature acquisition module is used for acquiring user features of a target user and object features of an object to be recommended, and respectively serving as the target user features and the target object features;

a recommendation prediction module, configured to input the target user feature and the target object feature into a pre-trained first recommendation prediction model, predict whether to recommend the object to be recommended to the target user, where the first recommendation prediction model is obtained by training in combination with a second recommendation prediction model, the first recommendation prediction model is configured to perform feature fusion on a feature of the user and a feature of the object included in input data to obtain a first fusion result, predict whether to recommend the object to be recommended to the target user based on the first fusion result, the second recommendation prediction model is configured to extract association features representing association relationships between features of the user and features of the object included in the input data, fuse the extracted association features to obtain a second fusion result, predict whether to recommend the object to the target user based on the second fusion result, and adjust a first prediction parameter based on a difference between the first fusion result and the second fusion result and a difference between the first prediction result output by the first recommendation prediction model and the second prediction result output by the second recommendation prediction model in a joint training process, so that the second prediction result approaches to the first model;

And the object recommending module is used for recommending the object to be recommended to the target user if the output result of the first recommendation prediction model indicates that the object to be recommended can be recommended to the target user.

In a fourth aspect of the embodiment of the present invention, there is also provided a model training apparatus, including:

the sample obtaining module is used for obtaining sample user characteristics of a sample user and sample object characteristics of a sample object;

the first recommendation prediction module is used for inputting the characteristics of the sample user and the characteristics of the sample object into a first initial model, carrying out characteristic fusion on the characteristics of the sample user and the characteristics of the sample object included in input data, and predicting whether to recommend the sample object to the sample user based on a first sample fusion result to obtain a first sample prediction result;

the second recommendation prediction module is used for inputting the sample user features and the sample object features into a second initial model, extracting sample association features between the user features and the object features included in the input data, fusing the extracted sample association features, and predicting whether to recommend the sample object to the sample user based on a second fusion result to obtain a second sample prediction result;

The model determining module is used for adjusting model parameters of the first initial model and the second initial model based on the difference between the first sample fusion result and the second sample fusion result and the difference between the first sample prediction result and the second sample prediction result until a preset training ending condition is met, so that the first sample fusion result approaches to the second sample fusion result, the first sample prediction result approaches to the second sample prediction result, the trained first initial model is determined to be a first recommendation prediction model for predicting whether to recommend an object to a user, and the trained second initial model is determined to be a second recommendation prediction model for predicting whether to recommend the object to the user.

In a fifth aspect of the embodiments of the present invention, there is also provided an electronic device, including a processor, a communication interface, a memory, and a communication bus, where the processor, the communication interface, and the memory complete communication with each other through the communication bus;

a memory for storing a computer program;

a processor for implementing the method steps of any of the above first aspects when executing a program stored on a memory.

In a sixth aspect of the embodiment of the present invention, there is provided another electronic device, including a processor, a communication interface, a memory, and a communication bus, where the processor, the communication interface, and the memory complete communication with each other through the communication bus;

A memory for storing a computer program;

a processor configured to implement the method steps of any of the second aspects when executing a program stored on a memory.

In a seventh aspect of the embodiments of the present invention, there is also provided a computer readable storage medium having stored therein a computer program which, when executed by a processor, implements the method steps of any of the first aspects.

In an eighth aspect of embodiments of the present invention, there is also provided another computer readable storage medium having stored therein a computer program which when executed by a processor implements the method steps of any of the second aspects.

In a further aspect of embodiments of the present invention, there is also provided a computer program product comprising instructions which, when run on a computer, cause the computer to perform the method steps of any of the first aspects described above.

In a further aspect of embodiments of the present invention, there is also provided a computer program product comprising instructions which, when run on a computer, cause the computer to perform the method steps of any of the second aspects described above.

According to the object recommendation method provided by the embodiment of the invention, the user characteristics of the user and the object characteristics of the object to be recommended are input into a pre-trained first recommendation prediction model, whether the object to be recommended is recommended to the user is predicted, and if the output result indicates that the object to be recommended can be recommended to the user, the object to be recommended is recommended to the user. Because the user has different interested degrees on different objects, the associated relation between the user and the different objects is different, the associated relation between the user and the object of interest of the user is different from the associated relation between the user and the object of no interest of the user, and the cross layer contained in the second recommendation prediction model can extract the associated relation between the user characteristics and the object characteristics, so that the second recommendation prediction model can predict whether to recommend the object to be recommended to the user according to the associated relation. Because the first recommendation prediction model is obtained by combined training with the second recommendation prediction model, the first prediction result of the first recommendation prediction model approaches to the second prediction result of the second recommendation prediction model, and therefore, the prediction result similar to the second recommendation prediction model can be obtained by using the first recommendation prediction model, and therefore, whether the object to be recommended is recommended to the user or not can be predicted by using the first recommendation prediction model, and therefore, the object can be quickly and accurately recommended to the user.

In addition, because the first recommendation prediction model does not contain a cross layer, the correlation features between the user features and the object features do not need to be extracted, and therefore compared with the second recommendation prediction model, the first recommendation prediction model has less calculation resources required to be consumed when predicting whether to recommend the object to be recommended to the user or not, and has higher calculation speed.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below.

Fig. 1 is a flow chart of an object recommendation method according to an embodiment of the present invention;

FIG. 2A is a flowchart of a first model training method according to an embodiment of the present invention;

FIG. 2B is a schematic structural diagram of a first initial model according to an embodiment of the present invention;

FIG. 3A is a flowchart of a second model training method according to an embodiment of the present invention;

FIG. 3B is a schematic structural diagram of a second initial model according to an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of a recommendation prediction model according to an embodiment of the present invention;

FIG. 5 is a flowchart of a third model training method according to an embodiment of the present invention;

FIG. 6 is a flowchart of a fourth model training method according to an embodiment of the present invention;

FIG. 7 is a flowchart of a fifth model training method according to an embodiment of the present invention;

fig. 8 is a schematic structural diagram of an object recommendation apparatus according to an embodiment of the present invention;

FIG. 9 is a schematic structural diagram of a first model training module according to an embodiment of the present invention;

FIG. 10 is a schematic structural diagram of a second model training module according to an embodiment of the present invention;

FIG. 11 is a schematic structural diagram of a first model training apparatus according to an embodiment of the present invention;

FIG. 12 is a schematic structural diagram of a second model training apparatus according to an embodiment of the present invention;

FIG. 13 is a schematic structural diagram of a third model training apparatus according to an embodiment of the present invention;

fig. 14 is a schematic structural diagram of an electronic device according to an embodiment of the present invention;

fig. 15 is a schematic structural diagram of another electronic device according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be described below with reference to the accompanying drawings in the embodiments of the present invention.

Because a great deal of time is required for a user to find an object of interest, the embodiment of the invention provides an object recommendation method and device for solving the problem.

In one embodiment of the present invention, there is provided an object recommendation method including:

inputting the target user characteristics and the target object characteristics into a pre-trained first recommendation prediction model, predicting whether to recommend the object to be recommended to the target user or not, wherein the first recommendation prediction model is obtained by jointly training with a second recommendation prediction model, the first recommendation prediction model is used for carrying out characteristic fusion on the characteristics of the user and the characteristics of the object included in input data to obtain a first fusion result and predicting whether to recommend the object to be recommended to the target user or not based on the first fusion result, the second recommendation prediction model is used for extracting association characteristics representing association relation between the characteristics of the user and the characteristics of the object included in the input data, carrying out fusion on the extracted association characteristics to obtain a second fusion result, predicting whether to recommend the object to be recommended to the target user or not based on the second fusion result, and adjusting parameters of the first recommendation prediction model based on difference between the first fusion result and the second fusion result output by the first recommendation prediction model in the joint training process so that the first recommendation prediction model approaches to the first fusion result and the second recommendation prediction result output by the second recommendation prediction model;

From the above, since the user has different interest degrees on different objects, the association relationship between the user and the different objects is different, the association relationship between the user and the object of interest is different from the association relationship between the user and the object of no interest, and the cross layer included in the second recommendation prediction model can extract the association relationship between the user feature and the object feature, so the second recommendation prediction model can predict whether to recommend the object to be recommended to the user according to the association relationship. Because the first recommendation prediction model is obtained by combined training with the second recommendation prediction model, the first prediction result of the first recommendation prediction model approaches to the second prediction result of the second recommendation prediction model, and therefore, the prediction result similar to the second recommendation prediction model can be obtained by using the first recommendation prediction model, and therefore, whether the object to be recommended is recommended to the user or not can be predicted by using the first recommendation prediction model, and therefore, the object can be quickly and accurately recommended to the user.

The method and the device for recommending the object provided by the embodiment of the invention are explained by a specific embodiment.

Referring to fig. 1, an embodiment of the present invention provides a flowchart of an object recommendation method, and specifically, the method includes the following steps S101 to S103.

S101: and obtaining the user characteristics of the target user and the object characteristics of the object to be recommended as the target user characteristics and the target object characteristics respectively.

Specifically, the object to be recommended may be video, audio, pictures, news, merchandise, and the like, and the scheme is not limited in particular.

The target object feature may be the case in the following (one) to (five).

In the case that the object to be recommended is a video, the target object features may include at least one of the following features:

video attribute features, such as duration of video, resolution of video, release time of video, and the like, and category features of categories to which video belongs, such as variety, movies, television shows, and the like.

(II) in the case that the object to be recommended is audio, the target object characteristics may include at least one of the following characteristics:

audio attribute features, such as duration of audio, time of release of audio, publisher of audio, etc., and category features of categories to which audio belongs, such as category features of categories to which audio belongs to popularity, rock, etc.

(III) in the case that the object to be recommended is a picture, the target object characteristics may include at least one of the following characteristics:

the attribute features of the picture, such as resolution of the picture, publisher of the picture, etc., and the category features of the category to which the picture belongs, such as category features of the category of animal pictures, landscape pictures, etc.

(IV) in the case that the object to be recommended is news, the target object characteristics may include at least one of the following characteristics:

the attribute features of news, such as the release time of the news, the number of characters contained in the news, the release of the news, and the like, and the category features of the category to which the news belongs, such as sports news, entertainment news, and the like.

(V) in the case that the object to be recommended is a commodity, the target object characteristics may include at least one of the following characteristics:

The attribute features of the commodity, such as price of the commodity, brand of the commodity, sales of the commodity and the like, and the category features of the category to which the commodity belongs, such as clothing, food and the like.

The target user feature is a feature representing a user attribute, and includes at least one of the following features:

user personal information characteristics such as the user's gender, age, academic, income, etc., object characteristics of objects viewed by the user.

Specifically, the object viewed by the user may be an object viewed in a preset time period, where the preset time period may be the time of creating the account for the user, or the time of last year, the time of last month, or the like.

S102: and inputting the target user characteristics and the target object characteristics into a pre-trained first recommendation prediction model, and predicting whether to recommend the object to be recommended to the target user.

Specifically, the first recommended prediction model is obtained by training in combination with the second recommended prediction model.

The first recommendation prediction model is used for carrying out feature fusion on features of a user and features of objects included in input data to obtain a first fusion result, and predicting whether to-be-recommended objects are recommended to a target user or not based on the first fusion result.

The first recommendation prediction model may perform feature fusion on features of a user and features of an object included in input data through a full-connection layer, where the full-connection layer may be a plurality of full-connection layers adjacent to each other, and perform feature fusion on the features multiple times.

In addition, whether to recommend an object to the user may be predicted by a regression function according to the first fusion result, so as to obtain a prediction result, for example, the regression function may be a softmax function.

In addition, the second recommendation prediction model is used for extracting association features between features of the user and features of the object included in the input data, fusing the extracted association features to obtain a second fusion result, and predicting whether to recommend the object to be recommended to the target user based on the second fusion result.

Specifically, the second recommendation prediction model may extract association features between features of the user and features of the object through a feature cross layer. The above-described association features may be used to represent an association between the target user feature and the target object feature.

The above-described associated features are not expressed in natural language, but are a type of computer-readable data.

For example, the association relationship between the personal information feature of the user and the category feature of the category to which the object belongs may be extracted, for example, the association relationship between the gender of the user and the category to which the video belongs, for example, the male user generally tends to watch the sports video, the association relationship between the male user and the sports video, in which the male user prefers to watch the sports video, may be generated through the feature cross layer, and the association feature may be represented in a vector form.

In addition, whether to recommend an object to the user may be predicted by a regression function according to the second fusion result, so as to obtain a prediction result, and for example, the regression function may be a softmax function.

In the process of carrying out combined training on the first recommended prediction model and the second recommended prediction model, the combined training is to respectively input the same sample data into an initial model of the first recommended prediction model and an initial model of the second recommended prediction model, and carry out parameter adjustment on the initial models together to carry out combined training.

In the combined training process, model parameters of the first recommended prediction model are adjusted based on the difference between the first fusion result and the second fusion result and the difference between the first prediction result output by the first recommended prediction model and the second prediction result output by the second recommended prediction model. So that the first fusion result approaches the second fusion result and the first prediction result approaches the second prediction result.

Since the parameters of the first recommended prediction model are adjusted according to the difference between the first prediction result and the second prediction result, it can be considered that the output information of the second recommended prediction model is referred to in the process of training to obtain the first recommended prediction model, that is, the first recommended prediction model is trained according to the second recommended prediction model, so that the prediction result of the first recommended prediction model approaches to the prediction result of the second recommended prediction model, and therefore the first recommended prediction model can be considered to be obtained by learning the second recommended prediction model, the first recommended prediction model can be referred to as a student model, and the second recommended prediction model can be referred to as a teacher model.

The first recommendation prediction model and the second recommendation prediction model are obtained through combined training, and the first recommendation prediction model and the second recommendation prediction model are subjected to parameter adjustment together, so that the first recommendation prediction model can refer to overall information between the first recommendation prediction model and the second recommendation prediction model in the training process. And under the condition that the second recommended prediction model is trained first and then the first recommended prediction model is trained, the training information in the training process of the first recommended prediction model and the training information in the training process of the second recommended prediction model are mutually independent. Because the whole information can be referred in the process of the combined training, the first prediction result output by the first recommended prediction model obtained through the combined training is more similar to the second prediction result output by the second recommended prediction model, and therefore the accuracy of the first recommended prediction model is higher.

Specifically, the first recommended prediction model and the second recommended prediction model may be obtained through training in steps S201 to S208, which will not be described in detail herein.

S103: and if the output result of the first recommendation prediction model indicates that the object to be recommended can be recommended to the target user, recommending the object to be recommended to the target user.

Specifically, the output result may be expressed in the form of a probability, for example, a probability that an object can be recommended to the user is 70%, a probability that an object cannot be recommended to the user is 30%, or the like, and if the probability that an object can be recommended to the user is greater than the probability that an object cannot be recommended to the user, it is indicated that an object to be recommended to the target user can be recommended.

The output result may be a predicted result of whether to recommend to the user.

Specifically, the object to be recommended may be sent to the login device of the target user, so that the login device of the target user may display the object to be recommended on the display interface, so as to push the object to be recommended to the target user.

The process of training to obtain the first recommended prediction model and the second recommended prediction model is described below with reference to the flowchart shown in fig. 2A and the model structure diagram shown in fig. 2B.

Referring to fig. 2A, an embodiment of the present invention provides a flow diagram of a first model training method, and referring to fig. 2B, an embodiment of the present invention provides a structural diagram of a first initial model.

Specifically, the method includes the following steps S201 to S208, where the first recommended prediction model and the second recommended prediction model are obtained through training in the following manner.

S201: sample user characteristics of the sample user and sample object characteristics of the sample object are obtained.

Specifically, the sample user feature corresponds to a target user feature, for example, in a case where the target user feature is a user personal information feature and an object feature of an object viewed by the user, the sample user feature is also the user personal information feature and the object feature of the object viewed by the user.

Similarly, the sample object corresponds to the target object, and the sample object feature corresponds to the target object feature, for example, when the target object is a video, the sample object is also a video, and when the target object feature is a video attribute feature and a category feature of a category to which the video belongs, the sample object feature is also a video attribute feature and a category feature of a category to which the video belongs.

S202: and inputting the sample user characteristics and the sample object characteristics into a first fusion layer of a first initial model, and fusing the sample user characteristics and the sample object characteristics to obtain a first sample fusion result.

The first initial model is an initial model of the first recommended prediction model, and the first initial model further includes a first discrimination layer. See fig. 2B for the first discrimination layer in the first initial model to the left.

Specifically, the first fusion layer may be formed by all connection layers adjacent to each other, after the first all connection layer receives the sample user feature and the sample object feature, the first all connection layer performs feature fusion on the sample user feature and the sample object feature, inputs a first sample fusion result output by the first all connection layer into the next all connection layer, continues feature fusion, and so on, inputs a first sample fusion result output by the last all connection layer into the first discrimination layer.

The first initial model may be manually modeled by a programmer, or may be a model which is already known in the art, and is not limited herein.

S203: and inputting the first sample fusion result into the first judging layer, and predicting whether to recommend the sample object to the sample user to obtain a first sample prediction result.

Specifically, the first discrimination layer may be a regression function, and the classification is performed according to the received first sample fusion result, and the regression function may be a softmax function.

S204: and inputting the sample user characteristics and the sample object characteristics into a cross layer of a second initial model, and extracting sample association characteristics between the sample user characteristics and the sample object characteristics.

Wherein the second initial model is an initial model of the second recommended prediction model, and the second recommended prediction model further includes: the second fusion layer and the second discrimination layer. See fig. 2B for a second fusion layer and a second discrimination layer in a second initial model to the right.

Specifically, the crossover layers may be second, third and higher order feature crossover layers, such as FM (Factorization Machine, factorizer), DNN (Deep Neural Networks, deep neural network), CIN (Compressed Interaction Network, compressed crossover network), transducer network, etc.

The second initial model may be obtained by manually modeling by a programmer, or may be obtained by a programmer in the prior art, and is not limited herein.

S205: and inputting the sample related features into the second fusion layer, and fusing the sample related features to obtain a second sample fusion result.

The second fusion layer may be composed of all connection layers adjacent to each other, and the same as the first fusion layer, after the first all connection layer in the second fusion layer receives the sample user feature and the sample object feature, the first all connection layer performs feature fusion on the sample user feature and the sample object feature, and inputs the second sample fusion result output by the first all connection layer into the next all connection layer, continues to perform feature fusion, and so on, inputs the second sample fusion result output by the last all connection layer into the second discrimination layer.

Moreover, as the second initial model can extract the sample correlation characteristics between the sample user characteristics and the sample object characteristics included in the input data, the process of extracting the sample correlation characteristics can improve the dimension of the input characteristics, and the first initial model directly performs the characteristic fusion processing on the input characteristics, so that the dimension of the input characteristics is lower. The process of feature fusion through the full-connection layer can reduce the number of dimensions of features, and the features with larger number of dimensions are subjected to feature fusion through more full-connection layers. The number of layers of fully connected layers contained in the second initial model is thus greater than the number of layers of fully connected layers in the first initial model.

S206: and inputting the second sample fusion result into the second judging layer, and predicting whether to recommend the sample object to the sample user or not to obtain a second sample prediction result.

Specifically, the second discriminating layer may be a regression function, and classified according to the received second sample fusion result, and the regression function may be a softmax function.

S207: and obtaining a first sample difference between the first sample fusion result and the second sample fusion result, and obtaining a second sample difference between the first sample prediction result and the second sample prediction result.

In one embodiment of the present invention, the mean square error between the first sample fusion result and the second sample fusion result with the same dimension may be calculated as the first sample difference.

In another embodiment of the present invention, the first sample difference may be calculated through steps M-N.

Step M: and predicting a result of fusion of the correlation features between the sample user features and the sample object features by adopting the first sample fusion result, and taking the result as a prediction fusion result.

Specifically, the first sample fusion result output by each full connection layer can be respectively input into a preset neural network to obtain the prediction fusion result.

Step N: and calculating the difference between the predicted fusion result and the second sample fusion result to be used as a first sample difference.

Specifically, the mean square error between the predicted fusion result and the second sample fusion result can be calculated and used as the first sample difference.

Because the second initial model comprises the cross layer, the correlation characteristic between the sample user characteristic and the sample object characteristic can be extracted, the correlation characteristic is input into the second fusion layer to enable the output second sample fusion result to be more accurate, but the first initial model does not comprise the cross layer, so that the difference between the first sample fusion result output by the first fusion layer and the second sample fusion result is larger, if the first sample difference between the first sample fusion result and the second sample fusion result is directly calculated, the parameters of the first initial model are adjusted according to the first sample difference, the adjustment amplitude is larger, the change of the first initial model in each training process is larger, and the accuracy of the first recommended prediction model obtained through training is lower. And predicting the result of fusion of the correlation features between the sample user features and the sample object features through the step M, wherein the difference between the result of fusion and the second sample fusion is smaller, and adjusting the parameters of the first initial model according to the first sample difference between the predicted fusion result and the second sample fusion result, so that the accuracy of the second recommended predicted model obtained through training is higher.

In addition, a mean square error between the first sample prediction result and the second sample prediction result may be calculated as the second sample difference.

S208: and adjusting model parameters of the first initial model and the second initial model based on the first sample difference and the second sample difference until a preset training ending condition is met, enabling the first sample fusion result to approach the second sample fusion result, enabling the first sample prediction result to approach the second sample prediction result, determining the trained first initial model as the first recommended prediction model, and determining the trained second initial model as the second recommended prediction model.

Specifically, the first sample difference and the second sample difference may be added as relative losses.

The first loss between the first sample predictor and the sample marker, and the second loss between the second sample predictor and the sample marker may be calculated with sample markers indicating whether the sample object is recommended to the sample user as training supervision information.

And (3) carrying out weighted calculation on the first loss, the second loss and the relative loss to obtain total loss, and adjusting model parameters of the first initial model and the second initial model according to the total loss.

In addition, the method does not adopt a gradient feedback technology in the model training process, namely gradient feedback blocking is performed, so that the influence of adjustment of parameters of the first initial model on the training of the second initial model can be prevented.

As can be seen from fig. 2B, the first initial model is shown on the left side and the second initial model is shown on the right side, and the first initial model does not include a cross layer compared to the second initial model. In a first initial model, user personal information characteristics, object characteristics of objects which are watched by a user, object attribute characteristics and class characteristics of classes which the objects belong to are input into a first fusion layer, then the first fusion layer inputs into a first judging layer, and a first sample prediction result is obtained through prediction of the first judging layer. In the second initial model, the personal information features of the user, the object features of the objects which are checked by the user, the object attribute features and the class features of the classes to which the objects belong are input into a cross layer, then the cross layer is input into a second fusion layer, then the second fusion layer is input into a second judging layer, and a second sample prediction result is obtained through prediction of the second judging layer. The dashed line indicates that the output results of the first fusion layer and the second fusion layer at the two ends of the dashed line are similar, and the first sample prediction result and the second sample prediction result at the two ends of the dashed line arrow are similar.

Therefore, between the first recommendation prediction model and the second recommendation prediction model which are obtained through training in the mode, more than the first prediction result is similar to the second prediction result, and the first fusion result is similar to the second fusion result, so that the prediction results of the first recommendation prediction model and the second recommendation prediction model are more similar, and the prediction results of the first recommendation prediction model are more accurate.

Referring to fig. 3A, a flow chart of a second model training method is provided in an embodiment of the present invention. Referring to fig. 3B, an embodiment of the present invention provides a schematic structural diagram of a second initial model. Specifically, in comparison with fig. 2B, the first initial model in fig. 3B further includes a first representation layer for adjusting the dimension of the feature, and the second initial model further includes a second representation layer for adjusting the dimension of the feature.

Compared to the embodiment shown in fig. 2A, before the step S202, the method further includes:

s209: and inputting the sample user characteristics and the sample object characteristics into the first representation layer, and adjusting the dimensions of the sample user characteristics and the sample object characteristics.

Specifically, the dimensions of the sample user features and the sample object features may be adjusted to the same dimensions.

For the discrete features in the sample user feature and the sample object feature, the discrete features may be adjusted to the same preset dimension by using an ebedding lookup table technology, for example, the preset dimensions may be 4, 6, and the like.

For continuous features in the sample user features and the sample object features, the continuous features can be adjusted to the same preset dimension through the fully connected network.

The above step S202 may be implemented by the following step S202A.

S202A: and inputting the sample user characteristics and the sample object characteristics after dimension adjustment into the first fusion layer, and fusing the sample user characteristics and the sample object characteristics to obtain a first sample fusion result.

Before the step S204, the method further includes:

s210: and inputting the sample user characteristics and the sample object characteristics into the second representation layer, and adjusting the dimensions of the sample user characteristics and the sample object characteristics.

In particular, the second presentation layer may be identical to the first presentation layer.

The above S204 can be realized by the following step S204A.

S204A: and inputting the sample user characteristics and the sample object characteristics after dimension adjustment into the cross layer, and extracting sample association characteristics between the sample user characteristics and the sample object characteristics.

As can be seen from fig. 3B, the first initial model is shown on the left side and the second initial model is shown on the right side, and the first initial model does not include a cross layer compared to the second initial model. In a first initial model, user personal information characteristics, object characteristics of objects which are watched by a user, object attribute characteristics and class characteristics of classes which the objects belong to are input into a first representation layer, then the first representation layer is input into a first fusion layer, then the first fusion layer is input into a first judging layer, and a first sample prediction result is obtained through prediction of the first judging layer. In the second initial model, the personal information features of the user, the object features of the objects which are watched by the user, the object attribute features and the class features of the classes to which the objects belong are input into a second representation layer, the second representation layer is input into a crossing layer, the crossing layer is input into a second fusion layer, the second fusion layer is input into a second judging layer, and a second sample prediction result is obtained through prediction of the second judging layer. The dashed line indicates that the output results of the first fusion layer and the second fusion layer at the two ends of the dashed line are similar, and the first sample prediction result and the second sample prediction result at the two ends of the dashed line arrow are similar.

From the above, the sample user features and the sample object features are respectively input into the first representation layer and the second representation layer, the sample user features and the sample object features are respectively adjusted to be in the same dimension, and are respectively input into the first fusion layer and the cross layer for further data processing.

Referring to fig. 4, an embodiment of the present invention provides a schematic structural diagram of a recommended prediction model.

As can be seen, the first recommended prediction model is on the left side and the second recommended prediction model is on the right side. The first recommended prediction model does not include a cross layer as compared to the second recommended prediction model, and the number of fully connected layers included in the first fusion layer is smaller than the number of fully connected layers included in the second fusion layer. In the first recommendation prediction model, user personal information characteristics, object characteristics of objects which are watched by a user, object attribute characteristics and category characteristics of categories to which the objects belong are input into a first representation layer, then the first representation layer is input into a first fusion layer, then the first fusion layer is input into a first judgment layer, and a first prediction result is obtained through prediction by the first judgment layer. In the second recommendation prediction model, the personal information features of the user, the object features of the objects which are watched by the user, the object attribute features and the class features of the classes to which the objects belong are input into a second representation layer, the second representation layer is input into a crossing layer, the crossing layer is input into a second fusion layer, the second fusion layer is input into a second judgment layer, and a second prediction result is predicted by the second judgment layer. The dashed line indicates that the output results of the full-connection layers at the two ends of the dashed line are similar, and the first prediction results at the two ends of the dashed line arrow are similar to the second prediction results.

Corresponding to the object recommendation method, the embodiment of the invention also provides a method for training the obtained model.

Referring to fig. 5, an embodiment of the present invention provides a flowchart of a third model training method, and specifically, the method includes the following steps S501-S504.

S501: sample user characteristics of the sample user and sample object characteristics of the sample object are obtained.

S502: and inputting the characteristics of the sample user and the characteristics of the sample object into a first initial model, carrying out characteristic fusion on the characteristics of the sample user and the characteristics of the sample object included in the input data, and predicting whether to recommend the sample object to the sample user based on a first sample fusion result to obtain a first sample prediction result.

Specifically, the step S502 may be implemented through steps S502A-S502B, which are not described in detail herein.

S503: and inputting the sample user features and the sample object features into a second initial model, extracting sample association features between the features of the user and the features of the object included in the input data, fusing the extracted sample association features, predicting whether to recommend the sample object to the sample user based on a second fusion result, and obtaining a second sample prediction result.

Specifically, the step S503 may be implemented by steps S503A to S503C, which will not be described in detail herein.

S504: and adjusting model parameters of the first initial model and the second initial model based on the difference between the first sample fusion result and the second sample fusion result and the difference between the first sample prediction result and the second sample prediction result until a preset training ending condition is met, so that the first sample fusion result approaches to the second sample fusion result, the first sample prediction result approaches to the second sample prediction result, determining the trained first initial model as a first recommended prediction model for predicting whether to recommend an object to a user, and determining the trained second initial model as a second recommended prediction model for predicting whether to recommend the object to the user.

Specifically, the step S504 may be implemented through the step O-step P:

step O: and predicting a result of fusion of the correlation between the sample user characteristics and the sample object characteristics by adopting the first sample fusion result, and taking the result as a prediction fusion result.

Step P: and the sample is marked as training supervision information, and model parameters of the first initial model and the second initial model are adjusted based on the difference between the prediction fusion result and the second sample fusion result and the difference between the first sample prediction result output by the first initial model and the second sample prediction result output by the second initial model.

Specifically, the above steps O-P are similar to the above steps M-N, and are not repeated here.

Steps S501 to S504 are similar to steps S201 to S208 shown in fig. 2A, and are not described herein.

Referring to fig. 6, an embodiment of the present invention provides a flowchart of a fourth model training method, and specifically, compared to the foregoing fig. 5, step S502 may be implemented through steps S502A-S502B.

S502A: and inputting the sample user characteristics and the sample object characteristics into a first fusion layer of the first initial model, and fusing the sample user characteristics and the sample object characteristics to obtain a first sample fusion result.

The first initial model further comprises a first judging layer.

S502B: and inputting the first sample fusion result into the first judging layer, and predicting whether to recommend the sample object to the sample user to obtain a first sample prediction result.

In comparison with the foregoing fig. 5, the above step S503 may be implemented by steps S503A to S503C:

S503A: and inputting the sample user features and the sample object features into a cross layer of the second initial model, and extracting sample association features between the sample user features and the sample object features.

Wherein the second initial model further comprises: the second fusion layer and the second discrimination layer.

S503B: and inputting the sample related features into the second fusion layer, and fusing the sample related features to obtain a second sample fusion result.

S503C: and inputting the second sample fusion result into the second judging layer, and predicting whether to recommend the sample object to the sample user or not to obtain a second sample prediction result.

Specifically, the steps S501 to S504 are similar to the steps S201 to S208 shown in fig. 2A, and are not described herein.

Referring to fig. 7, an embodiment of the present invention provides a flowchart of a fifth model training method, specifically, the first initial model further includes a first representation layer for adjusting the dimension of the feature, the second initial model further includes a second representation layer for adjusting the dimension of the feature, and, compared with the foregoing fig. 6,

before the step S502A, the method further includes:

S502C: and inputting the sample user characteristics and the sample object characteristics into the first representation layer, and adjusting the dimensions of the sample user characteristics and the sample object characteristics.

The step S502A described above can be implemented by the following step S502A 1.

S502A1: and inputting the sample user characteristics and the sample object characteristics after dimension adjustment into the first fusion layer, and fusing the sample user characteristics and the sample object characteristics to obtain a first sample fusion result.

Specifically, the steps S502C-S502A1 are similar to the steps S209-S202A shown in FIG. 3A, and are not repeated here.

Before the step S503A, the method further includes:

S503D: and inputting the sample user characteristics and the sample object characteristics into the second representation layer, and adjusting the dimensions of the sample user characteristics and the sample object characteristics.

The step S503A described above can be realized by the following step S503A 1.

S503A1: and inputting the sample user characteristics and the sample object characteristics after dimension adjustment into the cross layer, and extracting sample association characteristics between the sample user characteristics and the sample object characteristics.

Specifically, the steps S503D-S503A1 are similar to the steps S210-S204A shown in FIG. 3A, and are not repeated here.

Referring to fig. 8, an embodiment of the present invention provides a schematic structural diagram of an object recommendation apparatus, where the apparatus includes:

the feature acquisition module 801 is configured to obtain a user feature of a target user and an object feature of an object to be recommended, as a target user feature and a target object feature, respectively;

a recommendation prediction module 802, configured to input the target user feature and the target object feature into a pre-trained first recommendation prediction model, predict whether to recommend the object to be recommended to the target user, where the first recommendation prediction model is obtained by training in combination with a second recommendation prediction model, the first recommendation prediction model is configured to perform feature fusion on a feature of the user and a feature of the object included in input data to obtain a first fusion result, predict whether to recommend the object to be recommended to the target user based on the first fusion result, and the second recommendation prediction model is configured to extract association features representing association relationships between the feature of the user and the feature of the object included in the input data, fuse the extracted association features to obtain a second fusion result, predict whether to recommend the object to be recommended to the target user based on the second fusion result, and adjust a first parameter of the first recommendation prediction model to approach the first fusion result based on a difference between the first fusion result and the second fusion result and a difference between the first prediction result output by the first recommendation prediction model and the second prediction result output by the second recommendation prediction model in a joint training process;

And the object recommending module 803 is configured to recommend the object to be recommended to the target user if the output result of the first recommendation prediction model indicates that the object to be recommended can be recommended to the target user.

In an embodiment of the present invention, the object recommendation apparatus further includes a model training module 804, configured to train to obtain the first recommended prediction model and the second recommended prediction model, referring to fig. 9, a schematic structural diagram of the first model training module is provided, and specifically, the model training module 804 includes:

a sample obtaining submodule 804A for obtaining a sample user characteristic of a sample user and a sample object characteristic of a sample object;

a first feature fusion submodule 804B, configured to input the sample user feature and the sample object feature into a first fusion layer of a first initial model, and fuse the sample user feature and the sample object feature to obtain a first sample fusion result, where the first initial model is an initial model of the first recommended prediction model, and the first initial model further includes a first discrimination layer;

A first recommendation prediction submodule 804C, configured to input the first sample fusion result into the first discrimination layer, and predict whether to recommend the sample object to the sample user, so as to obtain a first sample prediction result;

the associated feature extraction submodule 804D is configured to input the sample user feature and the sample object feature into a cross layer of a second initial model, and extract a sample associated feature between the sample user feature and the sample object feature, where the second initial model is an initial model of the second recommended prediction model, and the second recommended prediction model further includes: a second fusion layer and a second discrimination layer;

a second feature fusion submodule 804E, configured to input the sample-related feature into the second fusion layer, and fuse the sample-related feature to obtain a second sample fusion result;

a second recommendation prediction submodule 804F, configured to input the second sample fusion result into the second discrimination layer, and predict whether to recommend the sample object to the sample user, so as to obtain a second sample prediction result;

a difference obtaining submodule 804G, configured to obtain a first sample difference between the first sample fusion result and a second sample fusion result, and obtain a second sample difference between the first sample prediction result and a second sample prediction result;

The model determining submodule 804H is configured to adjust model parameters of the first initial model and the second initial model based on the first sample difference and the second sample difference until a preset training end condition is met, so that the first sample fusion result approaches the second sample fusion result, the first sample prediction result approaches the second sample prediction result, determine the trained first initial model as the first recommended prediction model, and determine the trained second initial model as the second recommended prediction model.

Referring to fig. 10, an embodiment of the present invention provides a schematic structural diagram of a second model training module, specifically, the first initial model further includes a first representation layer for adjusting a dimension of a feature, the second initial model further includes a second representation layer for adjusting the dimension of the feature,

In comparison with the embodiment shown in fig. 9, the model training module 804 further includes:

a first dimension adjustment sub-module 804I, configured to input the sample user feature and the sample object feature into the first representation layer before the first feature fusion sub-module 804B inputs the sample user feature and the sample object feature into the first fusion layer of the first initial model, and adjust dimensions of the sample user feature and the sample object feature;

the first feature fusion submodule 804B is specifically configured to input the sample user feature and the sample object feature after the dimension adjustment into the first fusion layer;

a second dimension adjustment submodule 804J, configured to input the sample user feature and the sample object feature into the second representation layer before the associated feature extraction submodule 804D inputs the sample user feature and the sample object feature into the intersection layer of the second initial model, and adjust dimensions of the sample user feature and the sample object feature;

the associated feature extraction submodule 804D is specifically configured to input the sample user feature and the sample object feature after the dimension adjustment into the cross layer.

In one embodiment of the present invention, the difference obtaining submodule 804G is specifically configured to:

predicting a result of fusion of the correlation features between the sample user features and the sample object features by adopting a first sample fusion result, and taking the result as a prediction fusion result;

and calculating the difference between the prediction fusion result and the second sample fusion result to be used as a first sample difference, and obtaining a second sample difference between the first sample prediction result and the second sample prediction result.

From the above, since the second initial model includes the cross layer, the correlation feature between the sample user feature and the sample object feature can be extracted, and the correlation feature is input into the second fusion layer to make the output second sample fusion result more accurate, but the first initial model does not include the cross layer, so that the difference between the first sample fusion result output by the first fusion layer and the second sample fusion result is larger, if the first sample difference between the first sample fusion result and the second sample fusion result is directly calculated, the parameters of the first initial model are adjusted according to the first sample difference, the adjustment amplitude is larger, so that the first initial model changes more in the process of each training, and the accuracy of the trained first recommended prediction model is lower. And predicting the result of fusion of the correlation features between the sample user features and the sample object features through the step M, wherein the difference between the result of fusion and the second sample fusion is smaller, and adjusting the parameters of the first initial model according to the first sample difference between the predicted fusion result and the second sample fusion result, so that the accuracy of the second recommended predicted model obtained through training is higher.

Referring to fig. 11, an embodiment of the present invention provides a schematic structural diagram of a first model training apparatus, where the apparatus includes:

a sample obtaining module 1101, configured to obtain a sample user feature of a sample user and a sample object feature of a sample object;

the first recommendation prediction module 1102 is configured to input the sample user features and the sample object features into a first initial model, perform feature fusion on the features of the sample user and the features of the sample object included in the input data, and predict whether to recommend the sample object to the sample user based on a first sample fusion result, so as to obtain a first sample prediction result;

a second recommendation prediction module 1103, configured to input the sample user feature and the sample object feature into a second initial model, extract sample association features between the user feature and the object feature included in the input data, fuse the extracted sample association features, and predict whether to recommend the sample object to the sample user based on a second fusion result, so as to obtain a second sample prediction result;

the model determining module 1104 is configured to adjust model parameters of a first initial model and a second initial model based on a difference between the first sample fusion result and the second sample fusion result and a difference between the first sample prediction result and the second sample prediction result until a preset training end condition is satisfied, so that the first sample fusion result approaches the second sample fusion result, the first sample prediction result approaches the second sample prediction result, determine the trained first initial model as a first recommendation prediction model for predicting whether to recommend an object to a user, and determine the trained second initial model as a second recommendation prediction model for predicting whether to recommend an object to the user.

Referring to fig. 12, an embodiment of the present invention provides a schematic structural diagram of a second model training apparatus, and compared with the embodiment shown in fig. 11, the first recommendation prediction module 1102 includes:

a first feature fusion submodule 1102A, configured to input the sample user feature and the sample object feature into a first fusion layer of the first initial model, and fuse the sample user feature and the sample object feature to obtain a first sample fusion result, where the first initial model further includes a first discrimination layer;

the first recommendation prediction submodule 1102B is configured to input the first sample fusion result into the first discrimination layer, and predict whether to recommend the sample object to the sample user, so as to obtain a first sample prediction result;

The second recommendation prediction module 1103 includes:

an associated feature extraction submodule 1103A, configured to input the sample user feature and the sample object feature into a cross layer of the second initial model, and extract a sample associated feature between the sample user feature and the sample object feature, where the second initial model further includes: a second fusion layer and a second discrimination layer;

a second feature fusion submodule 1103B, configured to input the sample-related feature into the second fusion layer, and fuse the sample-related feature to obtain a second sample fusion result;

and a second recommendation prediction submodule 1103C, configured to input the second sample fusion result into the second discrimination layer, and predict whether to recommend the sample object to the sample user, so as to obtain a second sample prediction result.

Referring to fig. 13, an embodiment of the present invention provides a schematic structural diagram of a third model training apparatus, where, compared to the embodiment shown in fig. 12, the first initial model further includes a first representation layer for adjusting a feature dimension, and the second initial model further includes a second representation layer for adjusting a feature dimension;

the first recommendation prediction module 1102 further includes:

a first dimension adjustment submodule 1102C, configured to input the sample user feature and the sample object feature into the first representation layer before the first feature fusion submodule 1102A inputs the sample user feature and the sample object feature into the first fusion layer of the first initial model, and adjust dimensions of the sample user feature and the sample object feature;

the first feature fusion submodule 1102A is specifically configured to input the sample user feature and the sample object feature after the dimension adjustment into the first fusion layer;

the second recommendation prediction module 1103 further includes:

a second dimension adjustment sub-module 1103D, configured to input the sample user feature and the sample object feature into the second representation layer before the associated feature extraction sub-module 1103A inputs the sample user feature and the sample object feature into the intersection layer of the second initial model, and adjust dimensions of the sample user feature and the sample object feature;

The associated feature extraction submodule 1103A is specifically configured to input the sample user feature and the sample object feature after the dimension adjustment into the cross layer.

In one embodiment of the present invention, the model determining module 1104 is specifically configured to:

predicting a result of fusion of the correlation between the sample user characteristics and the sample object characteristics by adopting a first sample fusion result, and taking the result as a prediction fusion result;

and taking the sample mark as training supervision information, and adjusting model parameters of the first initial model and the second initial model based on the difference between the prediction fusion result and the second sample fusion result and the difference between the first sample prediction result output by the first initial model and the second sample prediction result output by the second initial model.

The embodiment of the present invention also provides an electronic device, as shown in fig. 14, including a processor 1401, a communication interface 1402, a memory 1403, and a communication bus 1404, where the processor 1401, the communication interface 1402, and the memory 1403 perform communication with each other through the communication bus 1404,

a memory 1403 for storing a computer program;

the processor 1401 is configured to implement the method steps described in any one of the object recommendation method embodiments when executing the program stored in the memory 1403.

When the electronic equipment provided by the embodiment of the invention is used for recommending the object, because the interested degree of the user on different objects is different, the incidence relation between the user and the object of interest and the incidence relation between the user and the object of no interest are different, and the incidence relation between the user characteristics and the object characteristics can be extracted by the crossed layer contained in the second recommendation prediction model, so that the second recommendation prediction model can predict whether to-be-recommended objects are recommended to the user according to the incidence relation. Because the first recommendation prediction model is obtained by combined training with the second recommendation prediction model, the first prediction result of the first recommendation prediction model approaches to the second prediction result of the second recommendation prediction model, and therefore, the prediction result similar to the second recommendation prediction model can be obtained by using the first recommendation prediction model, and therefore, whether the object to be recommended is recommended to the user or not can be predicted by using the first recommendation prediction model, and therefore, the object can be quickly and accurately recommended to the user.

The embodiment of the present invention further provides another electronic device, as shown in fig. 15, including a processor 1501, a communication interface 1502, a memory 1503, and a communication bus 1504, where the processor 1501, the communication interface 1502, and the memory 1503 complete communication with each other through the communication bus 1504,

a memory 1503 for storing a computer program;

the processor 1501 is configured to implement the method steps described in any of the model training method embodiments when executing the program stored in the memory 1503.

When the electronic equipment provided by the embodiment of the invention is used for model training, not only the first prediction result is similar to the second prediction result, but also the first fusion result is similar to the second fusion result between the first recommendation prediction model and the second recommendation prediction model obtained through training in the mode, so that the prediction results of the first recommendation prediction model and the second recommendation prediction model are more similar, and the prediction results of the first recommendation prediction model are more accurate.

The communication bus mentioned by the above terminal may be a peripheral component interconnect standard (Peripheral Component Interconnect, abbreviated as PCI) bus or an extended industry standard architecture (Extended Industry Standard Architecture, abbreviated as EISA) bus, etc. The communication bus may be classified as an address bus, a data bus, a control bus, or the like. For ease of illustration, the figures are shown with only one bold line, but not with only one bus or one type of bus.

The communication interface is used for communication between the terminal and other devices.

The memory may include random access memory (Random Access Memory, RAM) or non-volatile memory (non-volatile memory), such as at least one disk memory. Optionally, the memory may also be at least one memory device located remotely from the aforementioned processor.

The processor may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU for short), a network processor (Network Processor, NP for short), etc.; but also digital signal processors (Digital Signal Processor, DSP for short), application specific integrated circuits (Application Specific Integrated Circuit, ASIC for short), field-programmable gate arrays (Field-Programmable Gate Array, FPGA for short) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

In a further embodiment of the present invention, a computer readable storage medium is provided, in which a computer program is stored, which when executed by a processor, implements the method steps of any of the object recommendation method embodiments described in the previous embodiments.

When the computer program stored in the computer readable storage medium provided by the embodiment is used for recommending the object, because the interest degree of the user on different objects is different, the association relationship between the user and the objects of interest is different, and the association relationship between the user and the objects of interest and the association relationship between the user and the objects of no interest are different, the cross layer contained in the second recommendation prediction model can extract the association relationship between the user characteristics and the object characteristics, and therefore the second recommendation prediction model can predict whether to recommend the object to be recommended to the user according to the association relationship. Because the first recommendation prediction model is obtained by combined training with the second recommendation prediction model, the first prediction result of the first recommendation prediction model approaches to the second prediction result of the second recommendation prediction model, and therefore, the prediction result similar to the second recommendation prediction model can be obtained by using the first recommendation prediction model, and therefore, whether the object to be recommended is recommended to the user or not can be predicted by using the first recommendation prediction model, and therefore, the object can be quickly and accurately recommended to the user.

In yet another embodiment of the present invention, another computer readable storage medium is provided, where a computer program is stored, the computer program, when executed by a processor, implements the method steps of any of the model training method embodiments described in the previous embodiments.

When the computer program stored in the computer readable storage medium provided by the embodiment is used for model training, between the first recommendation prediction model and the second recommendation prediction model which are obtained through training in the mode, not only the first prediction result is similar to the second prediction result, but also the first fusion result is similar to the second fusion result, so that the prediction results of the first recommendation prediction model and the second recommendation prediction model are more similar, and the prediction results of the first recommendation prediction model are more accurate.

In a further embodiment of the present invention, a computer program product comprising instructions is also provided, which when run on a computer, causes the computer to perform the method steps of any of the object recommendation method embodiments described in the previous embodiments.

When the computer program product provided by the embodiment is executed to recommend the object, because the interested degree of the user on different objects is different, the incidence relation between the user and the object interested by the user is different from the incidence relation between the user and the object not interested by the user, and the incidence relation between the user characteristics and the object characteristics can be extracted by the crossed layer included in the second recommendation prediction model, so that the second recommendation prediction model can predict whether to recommend the object to be recommended to the user according to the incidence relation. Because the first recommendation prediction model is obtained by combined training with the second recommendation prediction model, the first prediction result of the first recommendation prediction model approaches to the second prediction result of the second recommendation prediction model, and therefore, the prediction result similar to the second recommendation prediction model can be obtained by using the first recommendation prediction model, and therefore, whether the object to be recommended is recommended to the user or not can be judged and predicted by the first recommendation prediction model, and therefore, the object can be recommended to the user quickly and accurately.

In yet another embodiment of the present invention, there is also provided another computer program product containing instructions that, when run on a computer, cause the computer to perform the method steps of any of the model training method embodiments described in the previous embodiments.

When the computer program product provided by the embodiment is executed to perform model training, between the first recommended prediction model and the second recommended prediction model obtained through training in the mode, more than the first prediction result is similar to the second prediction result, and the first fusion result is similar to the second fusion result, so that the prediction results of the first recommended prediction model and the second recommended prediction model are more similar, and the prediction results of the first recommended prediction model are more accurate.

In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, produces a flow or function in accordance with embodiments of the present invention, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another, for example, by wired (e.g., coaxial cable, optical fiber, digital Subscriber Line (DSL)), or wireless (e.g., infrared, wireless, microwave, etc.). The computer readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains an integration of one or more available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid State Disk (SSD)), etc.

It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

In this specification, each embodiment is described in a related manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for an apparatus, an electronic device, a computer-readable storage medium and a computer program product, the description is relatively simple, as it is substantially similar to the method embodiments, as relevant see the partial description of the method embodiments.

The foregoing description is only of the preferred embodiments of the present invention and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. An object recommendation method, the method comprising:

If the output result of the first recommendation prediction model indicates that the object to be recommended can be recommended to the target user, recommending the object to be recommended to the target user;

the first recommended prediction model and the second recommended prediction model are obtained through training in the following mode:

inputting the sample user features and the sample object features into a first fusion layer of a first initial model, and fusing the sample user features and the sample object features to obtain a first sample fusion result, wherein the first initial model is an initial model of the first recommended prediction model, and the first initial model further comprises a first judging layer;

inputting the first sample fusion result into the first judging layer, and predicting whether to recommend the sample object to the sample user to obtain a first sample prediction result;

inputting the sample user features and the sample object features into a cross layer of a second initial model, and extracting sample association features between the sample user features and the sample object features, wherein the second initial model is an initial model of the second recommended prediction model, and the second recommended prediction model further comprises: a second fusion layer and a second discrimination layer;

Inputting the sample association features into the second fusion layer, and fusing the sample association features to obtain a second sample fusion result;

inputting the second sample fusion result into the second judging layer, and predicting whether to recommend the sample object to the sample user to obtain a second sample prediction result;

obtaining a first sample difference between the first sample fusion result and a second sample fusion result, and obtaining a second sample difference between the first sample prediction result and a second sample prediction result;

and adjusting model parameters of the first initial model and the second initial model based on the first sample difference and the second sample difference until a preset training ending condition is met, enabling the first sample fusion result to approach the second sample fusion result, enabling the first sample prediction result to approach the second sample prediction result, determining the trained first initial model as the first recommended prediction model, and determining the trained second initial model as the second recommended prediction model.

2. The method of claim 1, wherein the step of determining the position of the substrate comprises,

the first initial model further comprises a first representation layer for adjusting the dimension of the feature, and the second initial model further comprises a second representation layer for adjusting the dimension of the feature;

Before the step of inputting the sample user features and the sample object features into the first fusion layer of the first initial model, the method further comprises:

inputting the sample user features and the sample object features into the first representation layer, and adjusting the dimensions of the sample user features and the sample object features;

the first fusion layer for inputting the sample user features and the sample object features into the first initial model comprises:

inputting the sample user characteristics and the sample object characteristics after dimension adjustment into the first fusion layer;

before the step of inputting the sample user features and the sample object features into the cross layer of the second initial model, the method further comprises:

inputting the sample user features and the sample object features into the second representation layer, and adjusting the dimensions of the sample user features and the sample object features;

the inputting the sample user features and sample object features into the intersection layer of the second initial model comprises:

and inputting the sample user characteristics and the sample object characteristics after the dimension adjustment into the cross layer.

3. The method of claim 1, wherein obtaining a first sample difference between the first sample fusion result and a second sample fusion result comprises:

and calculating the difference between the predicted fusion result and the second sample fusion result to be used as a first sample difference.

4. An object recommendation device, the device comprising:

The object recommending module is used for recommending the object to be recommended to the target user if the output result of the first recommendation prediction model indicates that the object to be recommended can be recommended to the target user;

wherein the apparatus further comprises: the model training module is used for training and obtaining the first recommendation prediction model and the second recommendation prediction model;

the model training module comprises:

the sample obtaining submodule is used for obtaining sample user characteristics of a sample user and sample object characteristics of a sample object;

the first feature fusion sub-module is used for inputting the sample user features and the sample object features into a first fusion layer of a first initial model, and fusing the sample user features and the sample object features to obtain a first sample fusion result, wherein the first initial model is an initial model of the first recommended prediction model, and the first initial model further comprises a first judging layer;

the first recommendation prediction sub-module is used for inputting the first sample fusion result into the first judging layer, predicting whether the sample object is recommended to the sample user or not, and obtaining a first sample prediction result;

The correlation feature extraction submodule is used for inputting the sample user features and the sample object features into a cross layer of a second initial model, and extracting sample correlation features between the sample user features and the sample object features, wherein the second initial model is an initial model of the second recommended prediction model, and the second recommended prediction model further comprises: a second fusion layer and a second discrimination layer;

the second feature fusion submodule is used for inputting the sample association features into the second fusion layer, and fusing the sample association features to obtain a second sample fusion result;

the second recommendation prediction sub-module is used for inputting the second sample fusion result into the second judging layer, predicting whether the sample object is recommended to the sample user or not, and obtaining a second sample prediction result;

the difference obtaining submodule is used for obtaining a first sample difference between the first sample fusion result and a second sample fusion result and obtaining a second sample difference between the first sample prediction result and a second sample prediction result;

the model determining submodule is used for adjusting model parameters of the first initial model and the second initial model based on the first sample difference and the second sample difference until a preset training ending condition is met, enabling the first sample fusion result to approach the second sample fusion result, enabling the first sample prediction result to approach the second sample prediction result, determining the trained first initial model as the first recommended prediction model, and determining the trained second initial model as the second recommended prediction model.

5. The apparatus of claim 4, wherein the first initial model further comprises a first representation layer for adjusting the dimensions of the feature, wherein the second initial model further comprises a second representation layer for adjusting the dimensions of the feature,

the model training module further comprises:

a first dimension adjustment sub-module, configured to input the sample user feature and the sample object feature into the first representation layer before the first feature fusion sub-module inputs the sample user feature and the sample object feature into a first fusion layer of a first initial model, and adjust dimensions of the sample user feature and the sample object feature;

the first feature fusion submodule is specifically configured to input the sample user feature and the sample object feature after dimension adjustment into the first fusion layer;

a second dimension adjustment sub-module, configured to input the sample user feature and the sample object feature into the second representation layer before the associated feature extraction sub-module inputs the sample user feature and the sample object feature into the intersection layer of the second initial model, and adjust dimensions of the sample user feature and the sample object feature;

The associated feature extraction sub-module is specifically configured to input the sample user feature and the sample object feature after the dimension adjustment into the cross layer.

6. The apparatus of claim 4, wherein the difference obtaining submodule is configured to:

7. The electronic equipment is characterized by comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are communicated with each other through the communication bus;

a memory for storing a computer program;

a processor for carrying out the method steps of any one of claims 1-3 when executing a program stored on a memory.

8. A computer-readable storage medium, characterized in that the computer-readable storage medium has stored therein a computer program which, when executed by a processor, implements the method steps of any of claims 1-3.