CN115630219A

CN115630219A - Training method and device of recommendation model and computer equipment

Info

Publication number: CN115630219A
Application number: CN202211100428.4A
Authority: CN
Inventors: 唐俪月
Original assignee: Industrial and Commercial Bank of China Ltd ICBC
Current assignee: Industrial and Commercial Bank of China Ltd ICBC
Priority date: 2022-09-09
Filing date: 2022-09-09
Publication date: 2023-01-20

Abstract

The application relates to a training method, a training device, computer equipment, a storage medium and a computer program product of a recommendation model, and relates to the technical field of big data. The method comprises the following steps: inputting the target object identification information into the embedding layer to obtain a target object identification vector; inputting the target object identification vector into an association network function to obtain a historical object identification vector associated with the target object; inputting the historical object identification vector into a deep learning model to obtain prediction information; determining a first evaluation loss function according to the real recommendation information and the prediction information of the target object, and determining a second evaluation loss function according to the real recommendation information and the prediction information of the historical object; and processing the first evaluation loss function and the second evaluation loss function by utilizing a regularization loss function and a minimum method, and updating the embedding layer and the associated network function to obtain the trained recommendation model. According to the method, the sparsity of the parameter solution of the recommendation model is improved on the basis that the accuracy of the regular term is not reduced, and the accuracy of the recommendation model is improved.

Description

Training method and device of recommendation model and computer equipment

Technical Field

The application relates to the technical field of big data, in particular to a training method and device for a recommendation model and computer equipment.

Background

In recent years, personalized recommendation systems have been widely used for various online services, including e-commerce, news media, and related websites. The recommendation system is a subset of the information filtering system, and aims to predict the possible behaviors of the user in the item according to the characteristic attributes of the user and the item, help the user to make a decision quickly, and improve the user satisfaction. In a formal online scene, new users and new projects are added continuously, and the cold start problem of recommending the new users and new projects becomes a research hotspot in two years.

However, in the prior art, when the recommendation model is used for recommending a new user and a new item, the cold start is usually faced with the problem that the prediction accuracy of the recommendation model is low.

Disclosure of Invention

In view of the above, it is necessary to provide a training method, an apparatus and a computer device for a recommendation model in order to solve the above technical problems.

In a first aspect, the present application provides a training method for a recommendation model. The method comprises the following steps:

inputting identification information of a target object into an embedding layer of a recommendation model to obtain an identification vector of the target object;

inputting the identification vector of the target object into an association network function of the recommendation model to obtain an identification vector of a historical object associated with the target object;

inputting the identification vector of the historical object into a deep learning model of the recommendation model to obtain the prediction information of the recommendation object corresponding to the target object;

determining a first evaluation loss function according to the real recommendation information and the prediction information of the target object, and determining a second evaluation loss function according to the real recommendation information and the prediction information of the historical object;

and updating the parameters of the embedding layer based on the first evaluation loss function, the regularization loss function and the minimum method, and updating the parameters of the correlation network function based on the second evaluation loss function, the regularization loss function and the minimum method to obtain a trained recommendation model.

In one embodiment, the updating the parameters of the embedding layer based on the first evaluation loss function, the regularization loss function, and the minimization method, and the updating the parameters of the association network function based on the second evaluation loss function, the regularization loss function, and the minimization method to obtain the trained recommendation model includes:

determining a target embedding layer parameter of the first evaluation loss function when the first evaluation loss function reaches a minimum value by utilizing the regularization loss function and a minimum method, and updating the parameter of the embedding layer to the target embedding layer parameter;

and determining a target associated network function parameter of the second evaluation loss function when the second evaluation loss function reaches the minimum value by utilizing the regularization loss function and the minimum method, and updating the parameter of the associated network function into the target associated network function parameter to obtain the trained recommendation model.

In one embodiment, the correlation network function includes a meta scaling network function and a meta offset network function;

the step of inputting the identification vector of the target object into the association network function of the recommendation model to obtain the identification vector of the historical object associated with the target object includes:

and after the identification vector of the target object is input into the meta-scaling network function, the output result of the meta-scaling network function is input into the meta-offset network function, so that the identification vector of the historical object associated with the target object is obtained.

In one embodiment, before the inputting the identification vector of the target object into the meta scaling network function and the inputting the output result of the meta scaling network function into the meta offset network function to obtain the identification vector of the history object associated with the target object, the method further includes:

inputting the characteristic information of the target object and the characteristic information of the historical object into an embedding layer of the recommendation model to obtain a characteristic vector of the target object and a characteristic vector of the historical object;

inputting the feature vector of the target object and the feature vector of the historical object into a meta-scaling network to obtain a meta-scaling network function;

inputting the identification information of the real recommended object of the target object and the identification information of the real recommended object of the historical object into an embedding layer of the recommendation model to obtain an identification vector of the target real recommended object and an identification vector of the historical real recommended object;

and inputting the identification vector of the target real recommended object and the identification vector of the historical real recommended object into the meta-offset network to obtain the meta-offset network function.

In one embodiment, the number of the identification information of the real recommended object of the target object is plural, the number of the identification information of the real recommended object of the history object is plural, and the inputting the identification vector of the target real recommended object and the identification vector of the history real recommended object into the meta-offset network to obtain the meta-offset network function includes:

inputting the identification vectors of the target real recommended objects into an average function to obtain a first average vector, and inputting the identification vectors of the historical real recommended objects into the average function to obtain a second average vector;

inputting the first average vector and the second average vector into the meta-migration network to obtain the meta-migration network function.

In one of the embodiments, the first and second parts of the device,

when the target object is a new user, the prediction information is used for indicating whether an interactive behavior exists between the new user and the historical item;

and in the case that the target object is a new item, the prediction information is used for indicating whether an interactive behavior exists between the new item and a historical user.

In a second aspect, the application also provides a recommendation method. The method comprises the following steps:

acquiring identification information of a target object;

inputting the identification information of the target object into a recommendation model to obtain the prediction information of the recommendation object corresponding to the target object;

wherein the recommendation model is obtained by training through the training method of the recommendation model according to the first aspect.

In a third aspect, the application further provides a training device for the recommendation model. The device comprises:

the identification vector acquisition module is used for inputting the identification information of the target object into the embedding layer of the recommendation model to obtain the identification vector of the target object;

the association conversion module is used for inputting the identification vector of the target object into an association network function of the recommendation model to obtain an identification vector of a historical object associated with the target object;

the deep learning module is used for inputting the identification vector of the historical object into a deep learning model of the recommendation model to obtain the prediction information of the recommendation object corresponding to the target object;

the evaluation loss module is used for determining a first evaluation loss function according to the real recommendation information and the prediction information of the target object and determining a second evaluation loss function according to the real recommendation information and the prediction information of the historical object;

and the updating module is used for updating the parameters of the embedding layer based on the first evaluation loss function, the regularization loss function and the minimum method, and updating the parameters of the correlation network function based on the second evaluation loss function, the regularization loss function and the minimum method so as to obtain a trained recommendation model.

In one embodiment, the update module is specifically configured to:

determining a target embedding layer parameter of the first evaluation loss function when the first evaluation loss function reaches the minimum value by utilizing the regularization loss function and a minimum method, and updating the parameter of the embedding layer to the target embedding layer parameter;

In one embodiment, the correlation network function comprises a meta scaling network function and a meta offset network function; the association conversion module is specifically configured to:

In one embodiment, the apparatus further comprises:

the characteristic vector acquisition module is used for inputting the characteristic information of the target object and the characteristic information of the historical object into an embedding layer of the recommendation model to obtain the characteristic vector of the target object and the characteristic vector of the historical object;

the meta scaling network function acquisition module is used for inputting the feature vector of the target object and the feature vector of the historical object into a meta scaling network so as to obtain a meta scaling network function;

the real recommended object vector acquisition module is used for inputting the identification information of the real recommended object of the target object into the embedding layer of the recommendation model to obtain the identification vector of the real recommended object;

and the meta-offset network function acquisition module is used for inputting the identification vector of the real recommended object into a meta-offset network so as to acquire the meta-offset network function.

In a fourth aspect, the application further provides a recommendation device. The device comprises:

the acquisition module is used for acquiring the identification information of the target object;

the recommendation module is used for inputting the identification information of the target object into a recommendation model to obtain the prediction information of the recommended object corresponding to the target object;

In a fifth aspect, the application further provides a computer device. The computer device comprises a memory storing a computer program and a processor implementing the steps of the first or second aspect when executing the computer program.

In a sixth aspect, the present application further provides a computer-readable storage medium. The computer readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the first aspect or the second aspect.

In a seventh aspect, the present application further provides a computer program product. The computer program product comprising a computer program that when executed by a processor performs the steps of the first aspect or the second aspect.

According to the training method, the training device, the computer equipment, the storage medium and the computer program product of the recommendation model, the identification information of the target object is input into the embedding layer of the recommendation model, and the identification vector of the target object is obtained; inputting the identification vector of the target object into an association network function of the recommendation model to obtain an identification vector of a historical object associated with the target object; inputting the identification vector of the historical object into a deep learning model of the recommendation model to obtain the prediction information of the recommendation object corresponding to the target object; determining a first evaluation loss function according to the real recommendation information and the prediction information of the target object, and determining a second evaluation loss function according to the real recommendation information and the prediction information of the historical object; and updating the parameters of the embedded layer based on the first evaluation loss function, the regularization loss function and the minimum method, and updating the parameters of the associated network function based on the second evaluation loss function, the regularization loss function and the minimum method to obtain a trained recommendation model. Therefore, the recommendation model can be trained by using a recommendation model training method, the first evaluation loss function and the second evaluation loss function are processed by using a regularization loss function and a minimum method, the embedding layer and the associated network function are updated, the sparsity of the solution of the recommendation model parameters can be improved on the basis of not reducing the accuracy by using the regularization item, and the accuracy of the recommendation model can be improved.

Drawings

FIG. 1 is a flow diagram of a method for training a recommendation model in one embodiment;

FIG. 2 is a block flow diagram of a method of training a recommendation model in another embodiment;

FIG. 3 is a flow diagram of a recommendation method in one embodiment;

FIG. 4 is a diagram of a training apparatus for a recommendation model in one embodiment;

FIG. 5 is a diagram of a recommender in one embodiment;

FIG. 6 is a diagram illustrating an internal structure of a computer device according to an embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

In an embodiment, as shown in fig. 1, the present application provides a method for training a recommendation model, and this embodiment is illustrated by applying the method to a terminal, it is to be understood that the method may also be applied to a server, and may also be applied to a system including a terminal and a server, and is implemented through interaction between the terminal and the server. The terminal can be but not limited to various personal computers, notebook computers, smart phones, tablet computers, internet of things equipment and portable wearable equipment, and the internet of things equipment can be smart sound boxes, smart televisions, smart air conditioners, smart vehicle-mounted equipment and the like. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, and the like. The server may be implemented as a stand-alone server or as a server cluster consisting of a plurality of servers. In this embodiment, the method includes the steps of:

step 101, inputting the identification information of the target object into the embedding layer of the recommendation model to obtain the identification vector of the target object.

The target object may be a cold user, and the target object may also be a cold item. The identification information of the target object may be a target object ID (IDentity).

In an embodiment of the application, the recommendation model may determine, according to the identification information of the cold user, the item prediction information corresponding to the cold user, so as to recommend the item for the cold user; the user prediction information corresponding to the cold item can also be predicted according to the identification information of the cold item, so that the user is recommended for the cold item. Therefore, in the process of training the recommendation model, the target object can be a cold user or a cold item. The identification information may be an ID of the cold user in case the target object is the cold user, and may be an ID of the cold item in case the target user is the cold item.

In an embodiment of the application, the embedding layer of the recommendation model may convert the input information into a vector, and the terminal may obtain the identification vector of the target object after inputting the identification information of the target object into the embedding layer. For example, the terminal may enter a cold user ID into the embedding layer resulting in a cold user ID identification vector u, and the terminal may also enter a cold item ID into the embedding layer resulting in a cold item ID identification vector v.

And 102, inputting the identification vector of the target object into an association network function of the recommendation model to obtain the identification vector of the historical object associated with the target object.

Wherein, in case the target object is a cold user, the history object associated with the cold user is a hot user; in the case where the target object is a cold item, the history object associated with the cold item is a hot item.

Similar characteristics exist between the target object and the historical object associated with the target object, so that a certain relation exists between the identification vector obtained by the identification information of the target object through the embedding layer and the identification vector obtained by the identification information of the historical object associated with the target object through the embedding layer.

In an embodiment of the application, the terminal converts the identification vector of the target object into the identification vector of the historical object associated with the target object by using an association network function, so that the conversion between cold users and hot users or the conversion between cold items and hot items is realized, and the deep learning model can predict the target object based on the identification vector of the historical object associated with the target object conveniently.

Specifically, when the target object is a cold user, the terminal inputs a cold user ID identification vector u output by the user's embedded layer to the association network function, and a hot user ID identification vector u may be obtained ^warm . Under the condition that the target object is a cold project, the terminal inputs a cold project ID identification vector v output by an embedding layer of the project into an association network function to obtain a hot project ID identification vector v ^warm 。

Step 103, inputting the identification vector of the historical object into the deep learning model of the recommendation model to obtain the prediction information of the recommendation object corresponding to the target object.

In one embodiment of the present application, the deep learning model may be a Data Communication Network (DCN), and the deep learning model may predict the target object based on information of a history object associated with the target object.

Specifically, the terminal may input the identification vector of the historical object into a deep learning model, and predict information output by the deep learning model, where the predict information includes predict data. In the case that the target object is a cold user, the deep learning model can predict the relation between the cold user and the item based on the relation between a hot user and the item which have similar characteristics with the cold user, and the prediction data output by the deep learning model can be used for representing whether click browsing behavior possibly exists between the cold user and the item; in the case that the target object is a cold item, the deep learning model can also predict the relationship between the cold item and the user based on the relationship between a hot item with similar characteristics to the cold item and the user, and the prediction data output by the deep learning model is also used for representing whether the cold item and the user may have click browsing behavior.

In an embodiment of the application, after step 101 and step 102, the terminal converts the identification vector of the cold user a into the identification vector of the hot user B and the identification vector of the hot user C, and the deep learning model may further calculate whether the click browsing behavior of the user a for the item may exist based on the conversion result, where if the click browsing behavior is predicted to exist, the prediction data in the prediction information is 1, and if the click browsing behavior is not predicted to exist, the prediction data in the prediction information is 0.

And 104, determining a first evaluation loss function according to the real recommendation information and the prediction information of the target object, and determining a second evaluation loss function according to the real recommendation information and the prediction information of the historical object.

Specifically, the first evaluation loss function includes a cold user evaluation loss function and a cold project evaluation loss function. The second evaluation loss function includes a hot user evaluation loss function and a hot item evaluation loss function. Under the condition that the target object is a cold user, the terminal can determine the evaluation loss function of the cold user according to the real recommendation information of the cold user and the prediction information output by the deep learning model, and the terminal can also determine the evaluation loss function of the hot user according to the real recommendation information of the hot user related to the cold user and the prediction information output by the deep learning model.

Under the condition that the target object is a cold item, the terminal can determine the cold item evaluation loss function according to the real recommendation information of the cold item and the prediction information output by the deep learning model, and the terminal can also determine the hot item evaluation loss function according to the real recommendation information of the hot item related to the cold item and the prediction information output by the deep learning model.

In one embodiment of the present application, the target object has real recommendation information y. When the target user is a cold user, the actual recommendation information of the target object indicates whether a click browsing behavior exists between the cold user and the project actually. When the target user is a cold item, the actual recommendation information of the target object indicates whether click browsing behavior exists between the cold item and the user actually.

The expression of the log-loss function is shown in (1):

wherein phi denotes a parameter of the embedding layerAnd y is the real recommendation information,

is the prediction information.

And 105, updating parameters of the embedding layer based on the first evaluation loss function, the regularization loss function and the minimum method, and updating parameters of the correlation network function based on the second evaluation loss function, the regularization loss function and the minimum method to obtain a trained recommendation model.

The recommendation model is applied to a recommendation system and used for recommending the target object.

Specifically, in the case that the target object is a cold user, the terminal may update the parameters of the embedded layer of the user based on the cold user evaluation loss function, the regularization loss function, and the minimum method, and may update the parameters of the associated network function of the user based on the hot user evaluation loss function, the regularization loss function, and the minimum method. When the target object is a cold project, the terminal can update parameters of an embedded layer of the project based on a cold project evaluation loss function, a regularization loss function and a minimum method, and update parameters of an associated network function of the project based on a hot project evaluation loss function, a regularization loss function and a minimum method. Among them, the Regularized loss function and The minimum method are FTRL (Follow The Regularized Leader) algorithm.

In an embodiment of the present application, before the identification information of the target object is input into the embedding layer of the recommendation model to obtain the identification vector of the target object, the method further includes:

collecting a sample data set, and carrying out model training on the recommendation model to obtain a trained recommendation model; the sample data set comprises target object information and historical object information, the target object information comprises identification information of a target object, characteristic information of the target object, identification information of a real recommended object of the target object and real recommended information of the target object, and the historical object information comprises identification information of a historical object, characteristic information of the historical object, identification information of the real recommended object of the historical object and real recommended information of the historical target object.

Specifically, before inputting the identification information of the target object into the embedded layer of the recommendation model to obtain the identification vector of the target object, the terminal needs to collect a sample data set, where the sample data set includes target object information and history object information, the target object information includes the identification information of the target object, the feature information of the target object, the identification information of the true recommendation object of the target object and the true recommendation information of the target object, and the history object information includes the identification information of the history object, the feature information of the history object, the identification information of the true recommendation object of the history object and the true recommendation information of the history target object.

Under the condition that the target object is a cold user, the sample data set acquired by the terminal can comprise a cold user ID, cold user characteristic information, item information of a cold user real recommended item, cold user real recommended information, a hot user ID, hot user characteristic information, identification information of a hot user real recommended user and hot user real recommended information, and the terminal can input the sample data set of the cold user into the recommendation model and finely adjust the recommendation model to realize model training of the recommendation model. Under the condition that the target object is a cold item, the sample data set acquired by the terminal can comprise cold item ID, cold item characteristic information, identification information of a cold item truly recommended item, identification information of a cold item truly recommended user, a hot item ID, hot item characteristic information, identification information of a hot item truly recommended user and hot item truly recommended information, and the terminal can input the sample data set of the cold item into the recommendation model, finely adjust the recommendation model and achieve model training of the recommendation model.

In the training method of the recommendation model, the identification information of the target object is input into the embedding layer of the recommendation model to obtain the identification vector of the target object; inputting the identification vector of the target object into an association network function of a recommendation model to obtain an identification vector of a historical object associated with the target object; inputting the identification vector of the historical object into a deep learning model of a recommendation model to obtain the prediction information of the recommendation object corresponding to the target object; determining a first evaluation loss function according to the real recommendation information and the prediction information of the target object, and determining a second evaluation loss function according to the real recommendation information and the prediction information of the historical object; and updating parameters of the embedded layer based on the first evaluation loss function, the regularization loss function and the minimum method, and updating parameters of the associated network function based on the second evaluation loss function, the regularization loss function and the minimum method to obtain a trained recommendation model. Therefore, the recommendation model can be trained by using a recommendation model training method, the first evaluation loss function and the second evaluation loss function are processed by using a regularization loss function and a minimum method, the embedding layer and the associated network function are updated, the sparsity of the solution of the recommendation model parameters can be improved on the basis of not reducing the accuracy by using the regularization term, and the accuracy of the recommendation model can be improved. In addition, the recommendation model training method can effectively improve the efficiency of the recommendation model in prediction, and enables the trained recommendation model to better resist complex fluctuation in online learning.

In an embodiment of the present application, the step 105 updates parameters of the embedded layer based on the first evaluation loss function, the regularization loss function, and the minimum method, and updates parameters of the associated network function based on the second evaluation loss function, the regularization loss function, and the minimum method, so as to obtain the trained recommendation model, specifically including: determining a target embedding layer parameter of the first evaluation loss function when the first evaluation loss function reaches the minimum value by utilizing a regularization loss function and a minimum method, and updating the parameter of the embedding layer into the target embedding layer parameter; and determining a target associated network function parameter of the second evaluation loss function when the second evaluation loss function reaches the minimum value by utilizing a regularization loss function and a minimum method, and updating the parameter of the associated network function into the target associated network function parameter to obtain a trained recommendation model.

Specifically, the first evaluation loss function includes a cold user evaluation loss function and a cold project evaluation loss function. The second evaluation loss function includes a hot user evaluation loss function and a hot item evaluation loss function.

When the target object is a cold user, the terminal can determine the minimum value of the evaluation loss function of the cold user by using a regularization loss function and a minimum method, and update the target embedding layer parameters of the user to the target embedding layer parameters. The terminal can also determine a target associated network function parameter of the user when the minimum value of the hot user evaluation loss function corresponding to the cold user is determined by utilizing a regularized loss function and a minimum method, and update the associated network function parameter of the user into the target associated network function parameter to finish training of a recommendation model of the cold user. When the target object is a cold project, the terminal can determine the minimum value of the cold project evaluation loss function by using a regularization loss function and a minimum method, and update the target embedding layer parameters of the project into the target embedding layer parameters. The terminal can also determine target associated network function parameters of the project when the minimum value of the hot project evaluation loss function corresponding to the cold project is determined by utilizing a regularization loss function and a minimum method, and update the parameters of the associated network function of the project into the target associated network function parameters to complete the training of the recommendation model of the cold project. The cold system comprises a cold user and a cold item, and the terminal completes the training of the cold user recommendation model and the cold item recommendation model when completing the training of the cold user recommendation model and the cold item recommendation model.

In one embodiment of the present application, the correlation network function includes a meta scaling network function and a meta offset network function; correspondingly, step 102 inputs the identification vector of the target object into the association network function of the recommendation model to obtain the identification vector of the historical object associated with the target object, which specifically includes: and after the identification vector of the target object is input into the meta-scaling network function, the output result of the meta-scaling network function is input into the meta-offset network function, so that the identification vector of the historical object associated with the target object is obtained.

Specifically, the terminal uses the meta-scaling network function to realize the conversion between the target object and the history object associated with the target object. In the recommendation process of the recommendation model, wrong interaction is often generated, such as false touch, unstable signal and the like. The noise generated by these situations can have a serious influence on the embedding process of the identification information of the target object in the embedding layer, and can also interfere with the process of converting the target object into a historical object associated with the target object. The terminal can reduce the influence of noise by using the meta-migration network function, and the terminal can make the recommendation model better resist the complex fluctuation existing in online learning by using the meta-migration network function.

When the target object is a cold user, the terminal may input the identification vector of the cold user into the meta-scaling network function of the user, and then input the output result of the meta-scaling network function of the user into the meta-offset network of the user, so as to obtain the identification vector of the hot user associated with the cold user. When the target object is a cold item, the terminal may input the identification vector of the cold item into the meta-scaling network function of the item, and then input the output result of the meta-scaling network function of the item into the meta-offset network of the item, so as to obtain the identification vector of the hot item associated with the cold item.

In an embodiment of the application, after the identification vector of the target object is input to the meta-scaling network function, before the output result of the meta-scaling network function is input to the meta-migration network function to obtain the identification vector of the history object associated with the target object, the method further includes:

inputting the characteristic information of the target object and the characteristic information of the historical object into an embedding layer of a recommendation model to obtain a characteristic vector of the target object and a characteristic vector of the historical object;

inputting the identification information of the real recommended object of the target object and the identification information of the real recommended object of the historical object into an embedding layer of a recommendation model to obtain an identification vector of the target real recommended object and an identification vector of the historical real recommended object;

and inputting the identification vector of the target real recommended object and the identification vector of the historical real recommended object into a meta-offset network to obtain a meta-offset network function.

Specifically, in the case that the target object is a cold user, the true recommendation object of the target object is an item having an interaction history with the cold user, and the true recommendation object of the history object is an item having an interaction history with the hot user. In the case that the target object is a cold item, the true recommendation object of the target object is a user who has an interaction history with the cold item, and the true recommendation object of the history object is a user who has an interaction history with the hot item.

The terminal can input the feature vector of the target object and the feature vector of the historical object into the meta-scaling network to obtain a meta-scaling network function, and convert the information of the target object into the information of the historical object associated with the target object. In addition, the terminal can input the identification information of the real recommended object of the target object and the identification information of the real recommended object of the history object into the metashift network to obtain the metashift network function. The terminal utilizes the real recommendation object information with interactive history as input information, so that noise interference can be effectively reduced, the stability of the recommendation model is improved, and the recommendation accuracy of the recommendation model is improved.

In an embodiment of the application, the number of the identification information of the real recommended object of the target object is multiple, the number of the identification information of the real recommended object of the history object is multiple, and the identification vector of the target real recommended object and the identification vector of the history real recommended object are input into the meta-migration network to obtain the meta-migration network function, including:

inputting the identification vectors of the plurality of target real recommended objects into an average function to obtain a first average vector, and inputting the identification vectors of the plurality of historical real recommended objects into the average function to obtain a second average vector;

the first mean vector and the second mean vector are input to a meta-migration network to obtain a meta-migration network function.

Specifically, the target object may have a plurality of real recommended objects, and the history object may also have a plurality of real recommended objects, so that the terminal inputs the identification information of the real recommended objects of the plurality of target objects into the embedding layer, and can obtain the identification vectors of the plurality of target real recommended objects. The terminal can process the identification vectors of the plurality of target real recommended objects by using the average function to obtain a first average vector, the terminal can input the identification information of the real recommended objects of the plurality of historical objects into the embedding layer to obtain the identification vectors of the plurality of historical real recommended objects, and then the terminal processes the identification vectors of the plurality of historical real recommended objects by using the average function to obtain a second average vector. The terminal may input the first average vector and the second average vector to the meta-migration network to obtain the meta-migration network function. The meta-bias network function obtained based on the identification information of the real recommended objects of the target objects and the identification information of the real recommended objects of the historical objects can enhance the noise immunity of the recommendation model, and improve the stability and accuracy of the recommendation model.

In an embodiment of the application, in the case that the target object is a new user, the prediction information is used for indicating whether an interactive behavior exists between the new user and the historical item;

in the case that the target object is a new item, the prediction information is used for indicating whether an interactive behavior exists between the new item and the historical user.

Specifically, in the case that the target object is a cold user (new user), the deep learning model may predict the relationship between the cold user and an item (history item) based on the relationship between a hot user (history user) and an item (history item) which have similar characteristics to the cold user, and the prediction data output by the deep learning model may be used to indicate whether there may be a click browsing behavior between the cold user and the item; in the case that the target object is a cold item (new item), the deep learning model can also predict the relationship between the cold item (new item) and the user (historical user) based on the relationship between a hot item (historical item) with similar characteristics to the cold item and the user (historical user), and the prediction data output by the deep learning model is also used for representing whether click browsing behaviors between the cold item and the user are possible.

In one embodiment of the present application, the terminal may be a two-category recommendation system for predicting a purchasing behavior or a click-to-browse behavior between a user and an item. The terminal can obtain a sample set, wherein each sample comprises a user, an item and a label y E {0,1}, the label y is used for representing purchasing behavior or click-to-browse behavior between the user and the item, and y is real recommendation information. The user and the project also contain some relevant features. For example, there are the following users: the cold user A, the hot user B and the hot user C have the following items: brand X (cold), brand Y (hot), brand Z (hot). Each user has user characteristics such as age, gender, occupation and income, and each commodity has project characteristics such as value, category and import or not. If a certain user such as A has a click browsing record between the X brand commodities, the y value of the relationship between the two is 1, otherwise, the y value is 0.

Referring to fig. 2, in an embodiment of the present application, the terminal may convert fewer features of the target object into a dense identification vector, i.e., an embedded vector, through the embedding layer. In the case where the target object is a cold user, the terminal will be able to enter the user's embedded layer with a cold user ID

Obtaining a cold user ID embedding vector (i.e., identification vector) u _j . The terminal can also input the user attribute (namely user characteristic information) into the embedded layer to obtain a user characteristic vector, and the expression of the user characteristic vector is

m represents the characteristic number of the user, the user attribute comprises a cold user characteristic attribute and a hot user characteristic attribute, and the user characteristic vector comprises a cold user characteristic vector and a hot user characteristic vector. The terminal can also input an item ID (namely identification information of a real recommended object of the target object) interacted with the user into the user embedding layer to obtain an embedding vector U (U) of the real recommended item _j ). In the case where the target object is a cold item, the terminal may input a cold item ID into an embedded layer of items

Get the Cold item ID embedding vector v _i . The terminal can input the project attribute (namely the project characteristic information) into the embedding layer to obtain a project characteristic vector, and the expression of the project characteristic vector is

n represents the feature number of the project, the project attributes comprise a cold project feature attribute and a hot project feature attribute, and the project feature vector comprises a cold project feature vector and a hot project feature vector. The terminal can also input the user ID interacted with the item (namely the identification information of the real recommended object of the target object) into the item embedding layer to obtain an embedding vector U (v) of the real recommended user _i )。

Similar features are often present for similar users or similar items, and thus there is a relationship between the cold ID insertion vector and the hot ID insertion vector. The terminal may utilize a meta-scaling network to convert the cold ID embedding vector to a better feature space to better fit the deep learning model. The feature conversion specifically comprises the following processes:

process 1:

and the terminal inputs the characteristic vector of the target object and the characteristic vector of the historical object into the element scaling network to obtain an element scaling network function.

Item meta scaling network function in case the target object is a cold item

The expression is as follows:

wherein, ω is _scale Is a parameter of h (-).

User meta-scaling network function in case the target object is a cold user

Is expressed as

Wherein, ω is _scale Is a parameter of h (-).

And (2) a process:

the meta-scaling network function can be viewed as a feature transformation that can convert a cold item ID to a hot item ID, and can also convert a cold item ID to a hot item ID.

In the case where the target object is a cold item, the output result of the meta-scaling network function outputs a hot item ID. The hot item ID expression at this time is:

in the case where the target object is a cold user, the output result of the meta scaling network function outputs a hot user ID. The hot user ID expression at this time is:

in the process of recommending the model, wrong interaction such as false touch, unstable signal and the like is often generated, and therefore, noise can seriously affect the cold ID embedding process. To mitigate the effects of noise, the terminal embeds the vector U (v) using the user ID interacted with the item _i ) And the item ID interacted with by the user is embedded into the vector U (U) _j ) The mean value of (a) is used as an input and input to the meta-migration network to obtain the function of the meta-migration network. The method for reducing the noise interference by using the metashift network function by the terminal specifically comprises the following processes:

process 1:

and the terminal inputs the identification vector of the real recommendation object into the meta-migration network to obtain a meta-migration network function. In the case where the target object is a cold item, the item element offset network function expression is as follows:

wherein w _shift Is a parameter of the g (·) function. Terminal utilization of simple unweighted mean function

Embedding vector U (v) into user ID interacted with item _i ) And (6) processing.

In the case where the target object is a cold user, the user meta-migration network function expression is as follows:

wherein w _shift Is a parameter of the g (·) function. Terminal using simple unweighted mean function

Embedding vector U (U) into item ID interacted with by user _j ) And (6) processing.

And (2) a process:

in the case where the target object is a cold item, the item element offsets the output result of the network function to output a hot item ID. The hot item ID expression at this time is:

in the case where the target object is a cold user, the output result of the user meta-migration network function outputs a hot user ID. The hot user ID expression at this time is:

after the terminal goes through the process, the identification vector of the target object is converted into a historical object identification vector associated with the target object, and then the terminal inputs the obtained historical object identification vector into a deep learning model to obtain prediction information, wherein the prediction information comprises prediction data. The terminal can judge whether click browsing behavior possibly exists between the user and the item according to the prediction information and the prediction data, and complete recommendation of the cold user or the cold item according to the prediction information. And the terminal completes the recommendation of the cold user and the cold item, namely completes the recommendation of the cold system.

The terminal can construct a cold user evaluation loss function and a cold project evaluation loss function according to the cold user embedded vector, the cold project embedded vector and the prediction data of the deep learning model

The terminal can also utilize the FTRL method to minimize

Respectively updating the user embedding layer and the item embedding layer

Similarly, the terminal can construct a hot user evaluation loss function corresponding to the cold user and a hot item evaluation loss function corresponding to the cold item according to the hot user embedding vector corresponding to the cold user, the hot item embedding vector corresponding to the cold item and the prediction data of the deep learning model

The terminal can also utilize the FTRL method to minimize

To update the user meta scaling network function separately

Item element scaling network function

User element migration network function and project element migration network function

To update by using FTRL method

The derivation of the parameters is for example, and other parameter updates are similarly available.

The parameter reduction is denoted as mu, its loss function

Simplified notation is l. t represents the t-th iteration, then the value of μ at the t +1 th iteration is:

the FTL adopts a calculation method of cumulative gradient, and the FTRL is formed by adding a regularization term on the basis of the FTL. In FTRL methods

Substitute for

Cumulative gradient of loss function for first round to t round, i.e.

And adding L1 regularization and L2 regularization, λ ₁ 、λ ₂ The coefficients of the respective L1 regular and L2 regular terms, the value of μ at the t +1 th iteration can be expressed as:

wherein, the first and the second end of the pipe are connected with each other,

then

Will be provided with

Unfolding to obtain:

is a constant that can be eliminated relative to mu to be optimized

Equation (13) can be simplified as:

the dimensions of the features are broken down into independent scalar minimization problems, i being the ith feature:

for which the second derivative method is used to obtain mu _t+1,i The update formula of (c) is:

for the case that the data of a certain feature of the cold user or the cold item in the recommendation model is less, the training rate corresponding to the feature dimension can be kept at a larger value alone without being in forced agreement with the training rates of other features. The FTRL algorithm has an independent learning rate for the features of each dimension, can better reflect the distribution of samples on different features, and can improve the sparsity by utilizing a regular term on the basis of not reducing the accuracy rate, thereby improving the accuracy and the efficiency of recommending the model.

The learning rate of the ith feature of the ttrl th round is as follows:

where α and β are hyper-parameters, β is generally 1, and α is selected as the case may be. g _s,i Is the partial derivative of the ith feature of the s-th round.

Then, the above equation (16) becomes:

in an embodiment of the present application, a recommendation method is further provided, as shown in fig. 3, a specific process includes the following steps:

step 301, acquiring identification information of a target object;

step 302, inputting the identification information of the target object into a recommendation model to obtain the prediction information of the recommendation object corresponding to the target object;

in the recommendation method, the recommendation model is obtained through training of the recommendation model training method, the target object is automatically and accurately recommended based on the obtained identification information of the target object, and the prediction information of the recommendation object corresponding to the target object is obtained, so that the user or the project is quickly, efficiently and accurately recommended.

It should be understood that, although the steps in the flowcharts related to the embodiments as described above are sequentially shown as indicated by arrows, the steps are not necessarily performed sequentially in the order indicated by the arrows. The steps are not limited to being performed in the exact order illustrated and, unless explicitly stated herein, may be performed in other orders. Moreover, at least a part of the steps in the flowcharts related to the above embodiments may include multiple steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of performing the steps or stages is not necessarily sequential, but may be performed alternately or alternately with other steps or at least a part of the steps or stages in other steps.

Based on the same inventive concept, the embodiment of the application also provides a training device of the recommendation model for realizing the training method of the recommendation model. The implementation scheme for solving the problem provided by the apparatus is similar to the implementation scheme described in the method, so the specific limitations in the following embodiments of the training apparatus for one or more recommendation models may refer to the limitations in the training method for the recommendation model, which are not described herein again.

In one embodiment of the present application, as shown in fig. 4, there is provided a training apparatus 400 for recommending a model, including: an identification vector acquisition module 410, an association conversion module 420, a deep learning module 430, an evaluation loss module 440, and an update module 450, wherein:

an identification vector obtaining module 410, configured to input identification information of the target object into an embedding layer of the recommendation model, to obtain an identification vector of the target object;

the association conversion module 420 is configured to input the identification vector of the target object into an association network function of the recommendation model to obtain an identification vector of a historical object associated with the target object;

the deep learning module 430 is configured to input the identification vector of the historical object into a deep learning model of the recommendation model to obtain prediction information of the recommendation object corresponding to the target object;

the evaluation loss module 440 is configured to determine a first evaluation loss function according to the actual recommendation information and the prediction information of the target object, and determine a second evaluation loss function according to the actual recommendation information and the prediction information of the historical object;

and the updating module 450 is configured to update the parameters of the embedded layer based on the first evaluation loss function, the regularization loss function, and the minimum method, and update the parameters of the associated network function based on the second evaluation loss function, the regularization loss function, and the minimum method, so as to obtain a trained recommendation model.

In an embodiment of the present application, the update module is specifically configured to:

determining a target embedding layer parameter of the first evaluation loss function when the first evaluation loss function reaches the minimum value by utilizing a regularization loss function and a minimum method, and updating the parameter of the embedding layer into the target embedding layer parameter;

and determining a target associated network function parameter of the second evaluation loss function when the second evaluation loss function reaches the minimum value by utilizing a regularization loss function and a minimum method, and updating the parameter of the associated network function into the target associated network function parameter to obtain a trained recommendation model.

In one embodiment of the present application, the correlation network function includes a meta scaling network function and a meta offset network function; the association conversion module is specifically configured to:

In one embodiment of the present application, the apparatus further comprises:

the element scaling network function acquisition module is used for inputting the characteristic vector of the target object and the characteristic vector of the historical object into an element scaling network to acquire an element scaling network function;

the real recommended object vector acquisition module is used for inputting the identification information of the real recommended object of the target object and the identification information of the real recommended object of the historical object into an embedding layer of the recommendation model to obtain the identification vector of the target real recommended object and the identification vector of the historical real recommended object;

and the meta-offset network function acquisition module is used for inputting the identification vector of the target real recommended object and the identification vector of the historical real recommended object into a meta-offset network so as to acquire a meta-offset network function.

In an embodiment of the application, the number of the identification information of the actual recommended object of the target object is multiple, the number of the identification information of the actual recommended object of the history object is multiple, and the meta-offset network function obtaining module is further configured to:

In one embodiment of the application, in the case that the target object is a new user, the prediction information is used for indicating whether an interactive behavior exists between the new user and the historical item;

Based on the same inventive concept, the embodiment of the present application further provides a recommendation device for implementing the above-mentioned related method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the method, so the specific limitations in one or more embodiments of the recommendation device provided below can be referred to the limitations on the recommendation method in the above, and details are not repeated here.

In one embodiment of the present application, as shown in fig. 5, there is provided a recommendation device 500, including: an acquisition module 510 and a recommendation module 520, wherein:

an obtaining module 510, configured to obtain identification information of a target object;

the recommending module 520 is configured to input the identification information of the target object into the recommending model to obtain prediction information of a recommended object corresponding to the target object;

and the recommendation model is obtained by training through the training method of the recommendation model.

The training device of the recommendation model and each module in the recommendation device may be wholly or partially implemented by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment of the present application, a computer device is provided, and the computer device may be a terminal, and the internal structure diagram thereof may be as shown in fig. 6. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program to run on the non-volatile storage medium. The communication interface of the computer device is used for communicating with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. The computer program is executed by a processor to implement a training method or a scale problem generation method for a scale problem generation model. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on the shell of the computer equipment, an external keyboard, a touch pad or a mouse and the like.

Those skilled in the art will appreciate that the architecture shown in fig. 6 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

In an embodiment of the present application, a computer device is provided, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method embodiments when executing the computer program.

In an embodiment of the application, a computer-readable storage medium is provided, on which a computer program is stored, which computer program, when being executed by a processor, carries out the steps of the above-mentioned method embodiments.

In an embodiment of the application, a computer program product is provided, comprising a computer program which, when being executed by a processor, carries out the steps of the above-mentioned method embodiments.

It should be noted that, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, presented data, etc.) referred to in the present application are information and data authorized by the user or sufficiently authorized by each party.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above may be implemented by hardware that is instructed by a computer program, and the computer program may be stored in a non-volatile computer-readable storage medium, and when executed, may include the processes of the embodiments of the methods described above. Any reference to memory, databases, or other media used in the embodiments provided herein can include at least one of non-volatile and volatile memory. The nonvolatile Memory may include a Read-Only Memory (ROM), a magnetic tape, a floppy disk, a flash Memory, an optical Memory, a high-density embedded nonvolatile Memory, a resistive Random Access Memory (ReRAM), a Magnetic Random Access Memory (MRAM), a Ferroelectric Random Access Memory (FRAM), a Phase Change Memory (PCM), a graphene Memory, and the like. Volatile Memory can include Random Access Memory (RAM), external cache Memory, and the like. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), among others. The databases referred to in various embodiments provided herein may include at least one of relational and non-relational databases. The non-relational database may include, but is not limited to, a block chain based distributed database, and the like. The processors referred to in the embodiments provided herein may be general purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing based data processing logic devices, etc., without limitation.

The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above examples only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present application. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A method of training a recommendation model, the method comprising:

2. The method for training the recommendation model according to claim 1, wherein the updating the parameters of the embedding layer based on the first evaluation loss function, the regularization loss function and the minimization method, and the updating the parameters of the association network function based on the second evaluation loss function, the regularization loss function and the minimization method to obtain the trained recommendation model comprises:

3. The training method of the recommendation model according to claim 1, wherein the correlation network function comprises a meta scaling network function and a meta offset network function;

4. The method for training a recommendation model according to claim 3, wherein before inputting the identification vector of the target object into the meta-scaling network function and inputting the output result of the meta-scaling network function into the meta-migration network function to obtain the identification vector of the history object associated with the target object, the method further comprises:

5. The method for training a recommendation model according to claim 4, wherein the number of identification information of the real recommendation object of the target object is plural, the number of identification information of the real recommendation object of the history object is plural, and the inputting the identification vector of the target real recommendation object and the identification vector of the history real recommendation object into the meta-offset network to obtain the meta-offset network function comprises:

6. The training method of a recommendation model according to claims 1 to 5,

when the target object is a new user, the prediction information is used for indicating whether an interactive behavior exists between the new user and a historical item;

7. A recommendation method, characterized in that the method comprises:

acquiring identification information of a target object;

wherein the recommendation model is obtained by training through the training method of the recommendation model according to any one of claims 1 to 6.

8. An apparatus for training a recommendation model, the apparatus comprising:

the association conversion module is used for inputting the identification vector of the target object into an association network function of the recommendation model to obtain the identification vector of the historical object associated with the target object;

and the updating module is used for updating the parameters of the embedded layer based on the first evaluation loss function, the regularization loss function and the minimum method, and updating the parameters of the associated network function based on the second evaluation loss function, the regularization loss function and the minimum method so as to obtain a trained recommendation model.

9. A recommendation device, characterized in that the device comprises:

10. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the method of any one of claims 1 to 6 or claim 7.

11. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 6 or claim 7.

12. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, realizes the steps of the method of any one of claims 1 to 6 or claim 7.