WO2024139666A1 - Training method and apparatus for dual-target domain recommendation model - Google Patents

Abstract

A training method and apparatus for a dual-target domain recommendation model. The method comprises: for each target domain, executing the following training operations: determining user representation data and item representation data according to data of interaction between a common user in a common user set corresponding to a dual-target domain and an item in an item set (201); converting the user representation data into another target domain to obtain converted user representation data (202); fusing the user representation data representing the same common user and the converted user representation data corresponding to another target domain to obtain fused user representation data (203); and performing training by using a machine learning method and by taking the fused user representation data and the item representation data as inputs and taking a tag indicating whether a common user and an item that are represented by the input data interact as an expected output, so as to obtain a dual-target domain recommendation model (204). The method achieves bidirectional data migration between dual-target domains, and improves the accuracy of the dual-target domain recommendation model.

Description

Training method and device for dual-target domain recommendation model

This disclosure claims priority to Chinese patent application No. 202211686837.7, filed on December 26, 2022, and entitled “Training method and device for dual-target domain recommendation model”, the entire contents of which are incorporated by reference in this disclosure.

Technical Field

The embodiments of the present disclosure relate to the field of artificial intelligence technology, specifically to deep learning technology, and more particularly to a training method and device for a dual-target domain recommendation model, and a dual-target domain recommendation method and device.

Background technique

Recommendation systems help alleviate the problem of information overload. Collaborative filtering methods are widely used techniques in modern recommendation systems. They can effectively learn embedded representations of users and items and make recommendations based on these embedded representations. However, these methods are plagued by the problem of data sparsity to a certain extent, making it difficult to accurately and effectively model user preferences, resulting in a sharp decline in the quality of recommendations. CDR (Cross Domain Recommendation) technology has recently been widely studied to alleviate the problem of data sparsity. It uses data in the auxiliary domain to improve the recommendation performance of the target domain. However, existing cross-domain recommendation methods only focus on one-way migration to improve the performance of the target domain with the help of auxiliary domain data, and there is a problem that only the target domain benefits.

Summary of the invention

The embodiments of the present disclosure provide a dual-target domain recommendation model training method and device, a dual-target domain recommendation method and device, a computer-readable medium, and an electronic device.

In a first aspect, an embodiment of the present disclosure provides a training method for a dual-target domain recommendation model, wherein for each target domain in the dual-target domains, the following training operations are performed: user representation data of the common users in the target domain and item representation data of the items in the item set are determined based on interaction data between the common users in the common user set corresponding to the dual-target domains and the items in the item set; the user representation data in the target domain is converted to another target domain to obtain converted user representation data corresponding to the target domain; The user representation data in the target domain representing the same shared user is fused with the converted user representation data corresponding to another target domain to obtain fused user representation data; a machine learning method is adopted, with the fused user representation data and item representation data as input, and a label representing whether there is interaction between the shared user corresponding to the input fused user representation data and the item corresponding to the input item representation data as the expected output, to train a dual-target domain recommendation model.

In some examples, the above-mentioned converting the user representation data in the target domain to another target domain to obtain the converted user representation data corresponding to the target domain includes: converting the user representation data in the target domain to another target domain through a trainable orthogonal mapping matrix to obtain the converted user representation data corresponding to the target domain.

In some examples, the above method also includes: after each iterative operation in the training process of the dual-target domain recommendation model, obtaining the orthogonal mapping matrix corresponding to the next iterative operation of the dual-target domain according to the orthogonal mapping matrix in the target domain and the orthogonal mapping matrix in another target domain.

In some examples, the above-mentioned converting the user representation data in the target domain to another target domain through a trainable orthogonal mapping matrix to obtain the converted user representation data corresponding to the target domain includes: converting the user representation data in the target domain to another target domain through a trainable orthogonal mapping matrix to obtain the initial converted user representation data corresponding to the target domain; using local differential privacy technology to process the initial converted user representation data to obtain the converted user representation data.

In some examples, the above-mentioned fusing the user representation data in the target domain representing the same shared user with the converted user representation data corresponding to another target domain to obtain the fused user representation data includes: determining a gated selection vector according to the similarity between the user representation data in the target domain representing the same shared user and the converted user representation data corresponding to another target domain through a trainable embedding fusion module; and fusing the user representation data in the target domain representing the same shared user with the converted user representation data corresponding to another target domain according to the gated selection vector to obtain the fused user representation data.

In some examples, the above gated selection vector is used to fuse the user representation data in the target domain and the converted user representation data corresponding to another target domain, thereby obtaining a fusion. The combined user representation data includes: performing weighted summation of the user representation data in the target domain representing the same common user and the converted user representation data corresponding to another target domain according to the gated selection vector to obtain the fused user representation data.

In some examples, the above-mentioned machine learning method uses fused user representation data and item representation data as input, and takes labels representing whether there is interaction between common users corresponding to the input fused user representation data and items corresponding to the input item representation data as expected output, to train a dual-target domain recommendation model, including: predicting the interaction probability between common users corresponding to the fused user representation data and items corresponding to the item representation data based on the fused user representation data and the item representation data; determining the loss between the interaction probability and the label through the objective function according to the interaction probability, the label and a preset normalization parameter; and training the dual-target domain recommendation model based on the loss.

In some examples, the above-mentioned interaction data is an interaction matrix that characterizes the interaction between common users in the common user set and projects in the project set; and the above-mentioned interaction data between common users in the common user set corresponding to the dual target domains and projects in the project set is used to determine the user representation data of common users and the project representation data of projects in the target domain, including: using matrix decomposition to determine the user representation data of common users and the project representation data of projects in the target domain based on the interaction matrix.

In a second aspect, an embodiment of the present disclosure provides a dual-target domain recommendation method, which performs the following recommendation operations for each target domain in the dual target domains: determining user representation data of the target common users and item representation data of the items in the item set based on the interaction data between the target common users corresponding to the dual target domains and the items in the item set; converting the user representation data in the target domain to another target domain to obtain converted user representation data corresponding to the target domain; fusing the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain fused user representation data; and determining the items to be recommended for the target common users based on the fused user representation data and item representation data through the trained dual-target domain recommendation model.

In some examples, the above-mentioned converting the user representation data in the target domain to another target domain to obtain the converted user representation data corresponding to the target domain includes: The trained orthogonal mapping matrix transforms the user representation data in the target domain into another target domain to obtain the transformed user representation data corresponding to the target domain.

In some examples, the above-mentioned orthogonal mapping matrix trained in the dual-target domain recommendation model is used to convert the user representation data in the target domain to another target domain to obtain the converted user representation data corresponding to the target domain, including: converting the user representation data in the target domain to another target domain through an orthogonal mapping matrix to obtain the initial converted user representation data corresponding to the target domain; using local differential privacy technology to process the initial converted user representation data to obtain the converted user representation data.

In some examples, the above-mentioned fusing the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain the fused user representation data includes: determining a gating selection vector according to the user representation data in the target domain and the converted user representation data corresponding to another target domain through an embedded fusion module trained in a dual-target domain recommendation model; and fusing the user representation data in the target domain with the converted user representation data corresponding to another target domain according to the gating selection vector to obtain the fused user representation data.

In some examples, the above-mentioned fusing the user representation data in the target domain with the converted user representation data corresponding to another target domain according to the gated selection vector to obtain the fused user representation data includes: performing weighted summation of the user representation data in the target domain and the converted user representation data corresponding to another target domain according to the gated selection vector to obtain the fused user representation data.

In some examples, the above-mentioned trained dual-target domain recommendation model determines the items to be recommended for the target common user based on the fused user representation data and item representation data, including: predicting the interaction probability between the target common user and the items represented by the item representation data based on the fused user representation data and item representation data; determining the items to be recommended for the target common user based on the interaction probability.

In some examples, the above-mentioned determining the user representation data of the target common users and the item representation data of the items in the item set according to the interaction data between the target common users corresponding to the dual target domains and the items in the item set includes: using a matrix decomposition method to determine the user representation data of the target common users and the item representation data of the items in the target domain based on the interaction matrix, wherein the interaction matrix represents the user representation data of the target common users and the item representation data of the items in the dual target domains. The interaction data between the common users in the common user set and the items in the item set corresponding to the domain.

In a third aspect, an embodiment of the present disclosure provides a training device for a dual-target domain recommendation model. For each target domain in the dual target domains, a training operation is performed by the following units: a first determination unit is configured to determine user representation data of common users and item representation data of items in the target domain based on interaction data between common users in a common user set corresponding to the dual target domains and items in an item set; a first obtaining unit is configured to convert the user representation data in the target domain to another target domain to obtain converted user representation data corresponding to the target domain; a second obtaining unit is configured to fuse the user representation data in the target domain representing the same common user with the converted user representation data corresponding to another target domain to obtain fused user representation data; a training unit is configured to adopt a machine learning method, with the fused user representation data and item representation data as input, and a label representing whether there is interaction between the common users corresponding to the input fused user representation data and the items corresponding to the input item representation data as the expected output, to train and obtain the dual-target domain recommendation model.

In some examples, the first obtaining unit is further configured to: convert the user representation data in the target domain to another target domain through a trainable orthogonal mapping matrix to obtain converted user representation data corresponding to the target domain.

In some examples, the above-mentioned device also includes: a matrix update unit, which is configured to obtain an orthogonal mapping matrix corresponding to the next iterative operation of the dual target domains according to the orthogonal mapping matrix in the target domain and the orthogonal mapping matrix in another target domain after each iterative operation in the training process of the dual target domain recommendation model.

In some examples, the first obtaining unit is further configured to: convert the user representation data in the target domain to another target domain through a trainable orthogonal mapping matrix to obtain the initial converted user representation data corresponding to the target domain; and use local differential privacy technology to process the initial converted user representation data to obtain converted user representation data.

In some examples, the second obtaining unit is further configured to: determine the gate selection vector according to the similarity between the user representation data in the target domain and the converted user representation data corresponding to another target domain that represent the same common user through a trainable embedding fusion module; The gated selection vector is used to fuse the user representation data in the target domain and the converted user representation data corresponding to another target domain, representing the same common user, to obtain fused user representation data.

In some examples, the second obtaining unit is further configured to: perform weighted summation of the user representation data in the target domain representing the same common user and the converted user representation data corresponding to another target domain according to the gated selection vector to obtain fused user representation data.

In some examples, the training unit is further configured to: predict the interaction probability between the common users corresponding to the fused user representation data and the items corresponding to the item representation data based on the fused user representation data and the item representation data; determine the loss between the interaction probability and the label based on the objective function, the interaction probability, the label and a preset normalization parameter; and train a dual-target domain recommendation model based on the loss.

In some examples, the interaction data is an interaction matrix that characterizes the interaction between common users in a common user set and projects in a project set; and the first determination unit is further configured to: determine the user representation data of common users and the project representation data of projects in the target domain based on the interaction matrix by matrix decomposition.

In a fourth aspect, an embodiment of the present disclosure provides a dual-target domain recommendation device, which performs a recommendation operation for each target domain in the dual target domains through the following units: a second determination unit is configured to determine the user representation data of the target common user and the item representation data of the item according to the interaction data between the target common user corresponding to the dual target domains and the items in the item set; a third obtaining unit is configured to convert the user representation data in the target domain to another target domain to obtain the converted user representation data corresponding to the target domain; a fourth obtaining unit is configured to fuse the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain the fused user representation data; a recommendation unit is configured to determine the items to be recommended of the target common user according to the fused user representation data and the item representation data through the trained dual-target domain recommendation model.

In some examples, the third obtaining unit is further configured to: convert the user representation data in the target domain to another target domain through the orthogonal mapping matrix trained in the dual-target domain recommendation model, and obtain the converted user representation data corresponding to the target domain.

In some examples, the third obtaining unit is further configured to: convert the user representation data in the target domain to another target domain through an orthogonal mapping matrix to obtain the initial converted user representation data corresponding to the target domain; and use local differential privacy technology to process the initial converted user representation data to obtain converted user representation data.

In some examples, the fourth obtaining unit is further configured to: determine a gating selection vector according to the user representation data in the target domain and the converted user representation data corresponding to another target domain through an embedded fusion module trained in the dual-target domain recommendation model; and fuse the user representation data in the target domain with the converted user representation data corresponding to another target domain according to the gating selection vector to obtain fused user representation data.

In some examples, the fourth obtaining unit is further configured to: perform weighted summation of the user representation data in the target domain and the converted user representation data corresponding to another target domain according to the gated selection vector to obtain fused user representation data.

In some examples, the recommendation unit is further configured to: predict the interaction probability between the target common user and the items represented by the item representation data based on the fused user representation data and item representation data; and determine the items to be recommended for the target common user based on the interaction probability.

In some examples, the second determination unit is further configured to: determine the user representation data of the target common users and the project representation data of the projects in the target domain based on the interaction matrix by matrix decomposition, wherein the interaction matrix represents the interaction data between the common users in the common user set and the projects in the project set corresponding to the dual target domains.

In a fifth aspect, an embodiment of the present disclosure provides a computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the method described in any one of the implementation modes of the first aspect and the second aspect is implemented.

In a sixth aspect, an embodiment of the present disclosure provides an electronic device, comprising: one or more processors; a storage device on which one or more programs are stored, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation manner of the first aspect and the second aspect.

BRIEF DESCRIPTION OF THE DRAWINGS

Other features, objects and advantages of the present disclosure will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

FIG1 is an exemplary system architecture diagram in which an embodiment of the present disclosure may be applied;

FIG2 is a flow chart of an embodiment of a training method for a dual-target domain recommendation model according to the present disclosure;

FIG3 is a schematic diagram of the structure of a dual-target domain recommendation model according to this embodiment;

FIG4 is a schematic diagram of an application scenario of the training method of the dual-target domain recommendation model according to this embodiment;

FIG5 is a flow chart of another embodiment of a training method for a dual-target domain recommendation model according to the present disclosure;

FIG6 is a flow chart of an embodiment of a dual-target domain recommendation method according to the present disclosure;

FIG7 is a flow chart of another embodiment of a dual-target domain recommendation method according to the present disclosure;

FIG8 is a structural diagram of an embodiment of a training device for a dual-target domain recommendation model according to the present disclosure;

FIG9 is a structural diagram of an embodiment of a dual-target domain recommendation device according to the present disclosure;

FIG. 10 is a schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present disclosure.

Detailed ways

The present disclosure is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other. The present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

It should also be noted that in the technical solution of this disclosure, the collection and The collection, updating, analysis, processing, use, transmission, storage and other aspects are in compliance with the relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good customs. Necessary measures are taken for user personal information to prevent illegal access to user personal information data and maintain the security of user personal information, network security and national security.

FIG. 1 shows an exemplary architecture 100 to which the dual-target domain recommendation model training method and apparatus, and the dual-target domain recommendation method and apparatus of the present disclosure can be applied.

As shown in FIG1 , the system architecture 100 may include terminal devices 101, 102, 103, a network 104, and a server 105. The communication connection between the terminal devices 101, 102, 103 constitutes a topological network, and the network 104 is used to provide a medium for a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Terminal devices 101, 102, 103 can be hardware devices or software that support network connection for data interaction and data processing. When terminal devices 101, 102, 103 are hardware, they can be various electronic devices that support network connection, information acquisition, interaction, display, processing and other functions, including but not limited to smart phones, tablet computers, e-book readers, laptop portable computers and desktop computers, etc. When terminal devices 101, 102, 103 are software, they can be installed in the electronic devices listed above. It can be implemented as multiple software or software modules for providing distributed services, for example, or it can be implemented as a single software or software module. No specific limitation is made here.

The server 105 may be a server that provides various services, for example, a background processing server that trains a dual-target domain recommendation model for both enterprise users corresponding to the dual target domains in response to receiving a model training request from the terminal devices 101, 102, and 103. For another example, a background processing server that recommends information to service users of both enterprise users through the dual-target domain recommendation models corresponding to the dual target domains. As an example, the server 105 may be a cloud server.

It should be noted that the server can be hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers or as a single server. When the server is software, it can be implemented as multiple software or software modules (such as software or software modules for providing distributed services), or it can be implemented as a single software or software module, which is not specifically limited here.

It should also be noted that the training method of the dual-target domain recommendation model and the dual-target domain recommendation method provided in the embodiments of the present disclosure can be executed by a server, or by a terminal device, or by a server and a terminal device in cooperation with each other. Accordingly, the various parts (e.g., various units) included in the training device of the dual-target domain recommendation model and the dual-target domain recommendation device can be all set in the server, or all set in the terminal device, or can be set in the server and the terminal device respectively.

It should be understood that the number of terminal devices, networks, and servers in FIG1 is merely illustrative. Depending on the implementation requirements, there may be any number of terminal devices, networks, and servers. When the training method of the dual-target domain recommendation model and the electronic device on which the dual-target domain recommendation method runs do not need to transmit data with other electronic devices, the system architecture may only include the training method of the dual-target domain recommendation model and the electronic device (e.g., server or terminal device) on which the dual-target domain recommendation method runs.

2, a flow chart 200 of an embodiment of a training method for a dual-target domain recommendation model is shown. For each target domain in the dual-target domain, a training operation represented by the following steps is performed:

Step 201 : determining user representation data of common users and item representation data of items in the target domain according to interaction data between common users in a common user set corresponding to dual target domains and items in an item set.

In this embodiment, the executing entity of the training method of the dual-target domain recommendation model (for example, the terminal device or server in Figure 1) can use a wired network connection method or a wireless network connection method to obtain the interaction data between the common users in the common user set corresponding to the dual target domains and the items in the item set remotely or locally; and determine the user representation data of the common users and the item representation data of the items in the target domain based on the interaction data.

The dual target domains may be two domains with a certain correlation. As an example, the dual target domains are the first target domain corresponding to the e-commerce platform and the second target domain corresponding to the short video platform. There is a certain correlation between the item information browsed on the e-commerce platform and the video information browsed by users on the short video platform. Both the e-commerce platform and the short video platform can use each other's data to solve their own data sparsity problems.

The common user set includes common users in the two target domains. Each target domain corresponds to a project set, and the project set includes projects in the corresponding target domain. Projects represent the type of information that users interact with in the target domain. Taking e-commerce platforms as an example, projects can be various items; taking short video platforms, projects can be short videos containing information of a given type.

For each target domain in the dual target domain, the same encoding method can be used to determine the unique identifier of the user it serves; then, by comparing the unique identifiers of the users in the dual target domains, all common users in the dual target domains are determined to obtain a common user set; finally, for each target domain, based on the common user set, the interaction data between the common users in the target domain and the items in the item set is determined. It should be noted that the interaction data includes both the relationship data between the user and the items that have interacted, and the relationship data between the user and the items that have not interacted.

After determining the interaction data, the execution subject may encode the interaction data using the encoding model to obtain user representation data of common users in each target domain, that is, the user's embedded representation information, and project representation data of projects, that is, the project's embedded representation information. As an example, the execution subject may encode the interaction data using an existing embedded representation model of the target (for example, a recommendation model obtained using a collaborative filtering method) to obtain user representation data and project representation data.

3, a structural diagram of the dual-target domain recommendation model is shown. The training process of the dual-target domain recommendation model 300 includes four stages: embedding representation, embedding conversion, embedding fusion, and parameter update and synchronization.

In some optional implementations of this embodiment, the interaction data is an interaction matrix that characterizes the interaction between the common users in the common user set and the items in the item set.

As an example, the dual target domains include domain A and domain B, which have the same common user set U but different item sets. The item sets in domain A and domain B are denoted as _VA and _VB , respectively. ^RA represents the interaction matrix between common users and items in domain A; correspondingly, ^RB represents the interaction matrix between common users and items in domain B. Interaction matrix between users and items.

In domain A, each common user i can be represented as the i-th row of the interaction matrix ^RA , i.e. Similarly, each item j can be represented as the jth column of the interaction matrix ^RA , i.e.

In this implementation, the execution subject may perform step 201 in the following manner: determine the user representation data of common users and the project representation data of projects in the target domain based on the interaction matrix by matrix decomposition.

As an example, through deep matrix factorization (DMF) and neural matrix factorization (NeuMF), user representation data and item representation data can be generated in each target domain. Taking the deep matrix factorization method as an example, in domain A, two multi-layer networks are used to map user i and item j to the latent space respectively, and a low-dimensional embedding vector is obtained:

Among them, σ represents the nonlinear activation function ReLU, k represents the dimension of user representation data and item representation data, and They respectively represent the weights of different layers in the multi-layer network corresponding to the shared users and the weights of different layers in the multi-layer network corresponding to the projects.

Similarly, referring to the above embedding representation process in domain A, it can be determined that there are user representation data of user i in domain B: and items z represent data and Among them, z∈V _B .

In this implementation, matrix decomposition is used to determine user representation data and item representation data in each target domain, thereby improving the efficiency and accuracy of the embedding representation process.

Step 202: convert the user representation data in the target domain into another target domain to obtain converted user representation data corresponding to the target domain.

In this embodiment, the execution subject may convert the user representation data in the target domain into another target domain to obtain converted user representation data corresponding to the target domain.

In order to better understand the user preferences of common users in one target domain, external user representation data from users in another target domain is used, and vice versa. In this embodiment, the user representation data of common users in one target domain needs to be converted for domain adaptation before being sent to another target domain.

As an example, the execution subject performs domain adaptation conversion on the user representation data through a trainable conversion network, wherein the parameters in the conversion network can be updated during the training process.

As another example, before the execution subject sends the converted user representation data to another target domain, the data may be encrypted in an encryption manner to improve the security of the data transmission process.

In some optional implementations of the present embodiment, the execution entity may perform step 202 as follows: convert the user representation data in the target domain to another target domain through a trainable orthogonal mapping matrix to obtain the converted user representation data corresponding to the target domain.

The orthogonal mapping matrix has the following two advantages: first, since the orthogonal transformation preserves the inner product of the vector, the orthogonal mapping matrix preserves the similarity of the user representation data before and after the transformation; second, the inverse matrix of the orthogonal matrix is equivalent to its transpose. In this way, the transformation of the user representation data between the two target domains can directly use its transpose to simplify the learning process and reduce the computational complexity.

In some optional implementations of the present embodiment, the above-mentioned execution entity may perform the above-mentioned data conversion process in the following manner: first, through a trainable orthogonal mapping matrix, the user representation data in the target domain is converted to another target domain to obtain the initial converted user representation data corresponding to the target domain; then, the local differential privacy technology is used to process the initial converted user representation data to obtain the converted user representation data.

In this implementation, local differential privacy technology is first used to further hide the true distribution of the transformed user representation data, and then the hidden transformed user representation data is migrated to another target domain.

The local differential privacy technology is as follows: For any local differential privacy function f, with domain Dom(f) and range Ran(f), for any input t, t′∈Dom(f), the output t*∈Ran(f) has:
P[f(t)＝t ^* ] ^≤eεP [f(t′)＝t ^* ]

Local differential privacy technology ensures that algorithm f satisfies local differential privacy by controlling the similarity of the output results of any two records, that is, the output is always t ^* , so that the eavesdropper cannot confirm whether the input is t or t ^* .

The use of local differential privacy technology can not only prevent external attackers from inference attacks, but also This is to prevent corporate partners corresponding to another target domain from inferring the real user representation data based on the orthogonal mapping matrix, thereby infringing on business privacy.

As an example, there are user representation data of user i in domain A: Migration to domain B can be expressed as Similarly, there are user representation data of user i in domain B Migration to domain A can be expressed as in:

denote the orthogonal mapping matrices in domain A and domain B respectively, and the two orthogonal mapping matrices always remain synchronized and equal. [·] ^T denotes the transpose of the matrix. La(0,λ) denotes Laplace noise with mean 0, and λ controls the strength of the Laplace noise. The larger λ is, the larger the noise is, and the better the local differential privacy protection is.

Step 203 , the user representation data in the target domain representing the same common user is fused with the converted user representation data corresponding to another target domain to obtain fused user representation data.

In this embodiment, the execution entity may fuse the user representation data in the target domain representing the same shared user with the converted user representation data corresponding to another target domain to obtain fused user representation data.

As an example, for each common user in each target domain, the execution entity may directly add the user representation data corresponding to the common user in the target domain and the converted user representation data corresponding to the common user in another target domain to obtain fused user representation data.

As another example, for each shared user in each target domain, the execution entity may perform weighted summation of the user representation data corresponding to the shared user in the target domain and the converted user representation data corresponding to the shared user in another target domain to obtain fused user representation data.

In some optional implementations of this embodiment, the execution subject may perform step 203 in the following manner:

First, through a trainable embedding fusion module, based on the similarity between the user representation data in the target domain and the converted user representation data corresponding to another target domain that represent the same common user, Determine a gate selection vector; second, according to the gate selection vector, fuse the user representation data in the target domain representing the same common user with the converted user representation data corresponding to another target domain to obtain fused user representation data.

Taking domain A as an example, using _L1 distance To represent the user data in domain A The transformed user representation data corresponding to domain B The similarity between:

It can be understood that the user in domain B represents data The transformed user representation data corresponding to domain A The similarity between them is expressed as:

Ideally, A feature dimension in and In terms of similarity at the feature level, the smaller the parameter value of the feature dimension, the more similar the two target domains are in this specific dimension. At the same time, in each feature dimension, the distance between the user representation data in the domain and the corresponding transformed user representation data in the external domain contributes differently to the target domain. Therefore, the embedding fusion module using a two-layer fully connected neural network is used to automatically learn and derive the distance between the user representation data in domain A and the corresponding transformed user representation data in domain A. The gate selection vector to control the information flow of each feature dimension; similarly, the information in domain B can also be learned and derived The gate selection vector in:

in, and denote the trainable weights and biases in the embedding fusion module in domain A, respectively. and They represent the trainable weights and biases in the embedding fusion module in domain B, respectively, and σ represents the nonlinear activation function ReLU.

Existing cross-domain recommendation methods usually align and fuse features based on the embedding representation within the entire domain and the cross-domain embedding representation. The fusion methods include splicing, maximum pooling, average pooling and other operations. However, not all features contained in the auxiliary domain are beneficial to the target domain. If there is also data sparsity in the auxiliary domain, the migrated embedding representation is not sufficient, or the quality of the embedding representation in the target domain is sufficient and no additional supplement is needed, negative transfer will occur. That is, in the process of transfer learning, The knowledge learned in the auxiliary domain has a negative impact on the learning in the target domain.

In this implementation, information fusion in cross-domain recommendation is explicitly considered, and a gated selection vector is derived to extract fine-grained signals that are highly correlated with the target domain at the feature level, which can effectively avoid the negative transfer problem and thus improve the accuracy of the fused user representation data.

In some optional implementations of this embodiment, the above-mentioned execution entity can perform the above-mentioned second step in the following manner: according to the gated selection vector, weighted summation is performed on the user representation data in the target domain representing the same shared user and the converted user representation data corresponding to another target domain to obtain fused user representation data.

As an example, the above execution entity can perform data fusion through the following formula:

in, represents the fused user representation data in domain A, Represents the fused user representation data in domain B.

In this implementation, based on the gated selection vector, a weighted sum is performed on the user representation data in the target domain and the converted user representation data corresponding to another target domain, thereby further improving the accuracy of the obtained fused user representation data.

Step 204, adopting a machine learning method, taking the fused user representation data and the item representation data as input, and taking a label representing whether there is interaction between the common users corresponding to the input fused user representation data and the items corresponding to the input item representation data as the expected output, to train a dual-target domain recommendation model.

In this embodiment, the above-mentioned execution entity can adopt a machine learning method, with the fused user representation data and the project representation data as input, and a label representing whether there is interaction between the common users corresponding to the input fused user representation data and the projects corresponding to the input project representation data as the expected output, to train a dual-target domain recommendation model.

As an example, the above execution subject can iteratively perform the following update operations: first, the fused user representation data and item representation data are input into the initial dual-target domain recommendation model to predict the fused user The interaction probability between the common users represented by the representation data and the items represented by the item representation data is calculated; then, according to the objective function, the loss between the interaction probability and the label is calculated, where the label represents whether there is an interaction between the common users corresponding to the input fused user representation data and the items corresponding to the input item representation data; then, the update gradient is calculated according to the loss; finally, the initial dual-target domain recommendation model is updated by using the stochastic gradient descent method. The initial dual-target domain recommendation model can adopt a deep learning model.

In response to reaching a preset end condition, a trained dual-target domain recommendation model is obtained, wherein the preset end condition is, for example, that the training time exceeds a preset time threshold, the number of training times exceeds a preset number threshold, and the training loss tends to converge.

In some optional implementations of this embodiment, the execution subject may perform step 204 in the following manner:

First, based on the fused user representation data and item representation data, the interaction probability between the common users corresponding to the fused user representation data and the items corresponding to the item representation data is predicted.

Taking domain A as an example, a simple and widely used inner product model is used to estimate the interaction probability between common user i and item j

Second, through the objective function, the loss between the interaction probability and the label is determined according to the interaction probability, label and preset normalization parameters.

As an example, the model is trained in two target domains using the following objective functions:

Among them, _yij represents the label, Y ⁺ represents the set of positive samples where the user and item interaction can be observed, ^Y- represents the set of negative samples where the user and item interaction has never been observed, and max(R) represents the normalization parameter, which is the maximum rating in the entire dataset. For explicit feedback, in a 5-star system, max(R) is 5, and different _yij has different effects on the loss function; for implicit feedback, max(R) is 1, and _yij is 0 or 1.

Third, based on the loss, a dual-target domain recommendation model is trained.

The parameters that can be updated in the dual-target domain recommendation model include parameters of the embedding representation network that determines the user representation data and the item representation data, parameters in the orthogonal mapping matrix, and parameters of the embedding fusion model.

In this implementation, a specific training method for the dual-target domain recommendation model is provided, which further improves the accuracy of the trained dual-target domain recommendation model.

In some optional implementations of this embodiment, the above-mentioned execution entity may also perform the following operations: after each iterative operation in the training process of the dual-target domain recommendation model, the orthogonal mapping matrix corresponding to the next iterative operation of the dual-target domain is obtained according to the orthogonal mapping matrix in the target domain and the orthogonal mapping matrix in the other target domain.

Taking domain A as an example, domain B directly sends its updated orthogonal mapping matrix X ^B ′ to domain A in plain text. Domain A averages X ^B ′ and the updated orthogonal mapping matrix X ^A ′ in the domain to obtain a new orthogonal mapping matrix X ^A . The same is true for domain B, which obtains a new orthogonal mapping matrix X ^B . This ensures that the orthogonal mapping matrices in the two domains are always the same in each iteration.

In this embodiment, federated learning is introduced into the cross-domain recommendation scenario. In the federated learning framework of this embodiment, the service user data of the enterprise organization corresponding to each target domain will never leave the local area, but will always be stored on the corresponding local device, such as a local database. The federated learning architecture is designed as a peer-to-peer network structure to further reduce the risk of privacy leakage, that is, there is no curious or malicious third party. The two target domains only communicate twice in one iteration to exchange information related to the model.

In the first communication, the migrated transformed user representation data is exchanged between the two domains. If the transformed user representation data is directly migrated, the other target domain may infer the real data of the user in the target domain (inference attack). To this end, local differential privacy technology is used to add noise that follows the Laplace distribution to hide the user's real representation data before transferring the transformed user representation data to another domain, which further enhances the privacy protection of user data. Because this not only prevents external attackers from intercepting the transmitted information, but also the other party's corporate partners cannot restore the original user representation data through the orthogonal mapping matrix. In local differential privacy, a higher noise intensity λ will lead to a decrease in data availability, thereby reducing recommendation performance. Therefore, there needs to be a balance between recommendation performance and privacy protection. Make trade-offs.

In the second communication, each domain exchanges updated orthogonal mapping matrices. In this embodiment, no privacy protection method is used for this communication process. Because the orthogonal mapping matrix represents the migration correspondence between the two domains and does not involve sensitive data. In addition, even if an external attacker intercepts the updated orthogonal mapping matrix and the user representation data after differential privacy, it is difficult to infer valid information.

Continuing to refer to FIG. 4 , FIG. 4 is a schematic diagram 400 of an application scenario of the training method of the dual-target domain recommendation model according to the present embodiment. In the application scenario of FIG. 4 , first, in the target domain A, according to the interaction data 403 between the common users in the common user set 401 corresponding to the dual target domains A and B and the items in the item set 402, the user representation data 404 of the common users in the target domain A and the item representation data 405 of the items are determined; in the target domain B, according to the interaction data 407 between the common users in the common user set 401 corresponding to the dual target domains A and B and the items in the item set 406, the user representation data 408 of the common users in the target domain B and the item representation data 409 of the items are determined.

Then, the user representation data 404 in the target domain A is converted to the target domain B to obtain the converted user representation data 410 corresponding to the target domain A; the user representation data 408 in the target domain B is converted to the target domain A to obtain the converted user representation data 411 corresponding to the target domain B. Then, in the target domain A, the user representation data 404 representing the same common user is fused with the converted user representation data 411 corresponding to the target domain B to obtain the fused user representation data 412; in the target domain B, the user representation data 408 representing the same common user is fused with the converted user representation data 410 corresponding to the target domain A to obtain the fused user representation data 413.

Finally, for the dual target domains A and B, machine learning methods are used, with the fused user representation data and item representation data as input, and the labels representing whether there is interaction between the common users corresponding to the input fused user representation data and the items corresponding to the input item representation data as the expected output, to train and obtain the dual target domain recommendation models 414 and 415 corresponding to the domains A and B respectively.

The method provided by the above embodiment of the present disclosure performs bidirectional data migration between two target domains. The data sparsity problem in each target domain is solved, so that the enterprise users of both target domains can benefit from the data migration process; moreover, it is more in line with the actual situation under the dual target domain situation, which is conducive to the active participation of enterprise users of both sides of the dual target domains; in addition, based on the data migration process between the dual target domains, the accuracy of the recommendation results is improved.

Continuing to refer to FIG. 5 , a schematic process 500 of another embodiment of a training method for a dual-target domain recommendation model according to the present disclosure is shown. For each target domain in the dual-target domain, a training operation including the following steps is performed:

Step 501, through a trainable embedding representation module, according to the interaction data between the common users in the common user set corresponding to the dual target domains and the projects in the project set, determine the user representation data of the common users in the target domain and the project representation data of the projects.

Step 502: transform the user representation data in the target domain into another target domain through a trainable orthogonal mapping matrix to obtain the initial transformed user representation data corresponding to the target domain.

Step 503: Use local differential privacy technology to process the initial converted user representation data to obtain converted user representation data.

Step 504 , determining a gated selection vector through a trainable embedding fusion module according to the similarity between the user representation data in the target domain and the converted user representation data corresponding to another target domain that represent the same common user.

Step 505 , according to the gated selection vector, weighted sum is performed on the user representation data in the target domain and the converted user representation data corresponding to another target domain that represent the same shared user, to obtain fused user representation data.

Step 506 , predicting the interaction probability between the common users corresponding to the fused user representation data and the items corresponding to the item representation data based on the fused user representation data and the item representation data.

Step 507, determining the loss between the interaction probability and the label through the objective function according to the interaction probability, the label and the preset normalization parameter.

Step 508: Update the embedding representation module, orthogonal mapping matrix, and embedding fusion module according to the loss. The dual-target domain recommendation model is trained.

It can be seen from this embodiment that, compared with the embodiment corresponding to Figure 2, the process 500 of the training method of the dual-target domain recommendation model in this embodiment specifically illustrates the conversion process, hiding process, and fusion process of the user representation data. While ensuring the bidirectional migration of the dual target domains and the accuracy of the recommendation results, the security of the data in the dual target domains is further improved.

Continuing to refer to FIG. 6 , a process 600 of an embodiment of a dual-target domain recommendation method is shown. For each target domain in the dual-target domain, a recommendation operation including the following steps is performed:

Step 601 : determining user representation data of the target common users and item representation data of the items according to interaction data between the target common users and the items in the item set corresponding to the dual target domains.

In this embodiment, the executing entity of the dual-target domain recommendation method (for example, the terminal device or server in Figure 1) can determine the user representation data of the target common users and the project representation data of the projects based on the interaction data between the target common users corresponding to the dual target domains and the projects in the project set.

The dual target domains can be two fields with certain correlation. As an example, the dual target domains are the first target domain corresponding to the e-commerce platform and the second target domain corresponding to the short video platform. The project represents the type of information involved by the user in the target domain. Taking the e-commerce platform as an example, the project can be a variety of items; taking the short video platform, the project can be a short video containing a given type of information.

In this embodiment, after determining the target shared user, the above-mentioned execution entity can determine the unique identifier of the target shared user; then, based on the unique identifier, determine the interaction data between the target shared user and the projects in the project set; finally, by using the currently existing embedding representation module of the deep neural network, the interaction data is processed to obtain the user representation data of the target shared user and the project representation data of the project.

In some optional implementations of this embodiment, the execution subject may perform step 601 in the following manner: determine the user representation data of the target common users and the project representation data of the projects in the target domain based on the interaction matrix by matrix decomposition, wherein the interaction matrix represents the interaction between the common users in the common user set and the projects in the project set corresponding to the dual target domains. data.

Step 602: convert the user representation data in the target domain into another target domain to obtain converted user representation data corresponding to the target domain.

In this embodiment, the execution subject may refer to the implementation method in step 202 in embodiment 200 to convert the user representation data in the target domain into another target domain to obtain the converted user representation data corresponding to the target domain.

In some optional implementations of the present embodiment, the execution entity may perform step 602 as follows: convert the user representation data in the target domain to another target domain through the orthogonal mapping matrix trained in the dual-target domain recommendation model, and obtain the converted user representation data corresponding to the target domain.

In some optional implementations of the present embodiment, the above-mentioned execution entity may perform the above-mentioned data conversion process in the following manner: first, the user representation data in the target domain is converted to another target domain through an orthogonal mapping matrix to obtain the initial converted user representation data corresponding to the target domain; then, the local differential privacy technology is used to process the initial converted user representation data to obtain the converted user representation data.

Step 603: Fusing the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain fused user representation data.

In this embodiment, the execution subject may refer to the implementation method in step 203 in embodiment 200 to fuse the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain fused user representation data.

In some optional implementations of the present embodiment, the execution subject may perform step 603 in the following manner: first, the trained embedding fusion module in the dual-target domain recommendation model determines a gating selection vector based on the user representation data in the target domain and the converted user representation data corresponding to the other target domain; second, based on the gating selection vector, the user representation data in the target domain and the converted user representation data corresponding to the other target domain are fused to obtain fused user representation data.

In some optional implementations of this embodiment, the above-mentioned execution subject can be implemented in the following ways: Execute the second step: perform weighted summation of the user representation data in the target domain and the converted user representation data corresponding to another target domain according to the gated selection vector to obtain fused user representation data.

Step 604 , using the trained dual-target domain recommendation model, according to the fused user representation data and item representation data, determine the items to be recommended for the target common user.

In this embodiment, the execution subject can determine the recommended items for the target common user according to the fused user representation data and item representation data through the trained dual-target domain recommendation model. The dual-target domain recommendation model can be obtained through training in the above embodiments 200 and 500.

In some optional implementations of the present embodiment, the execution entity may perform step 604 in the following manner: first, predicting the interaction probability between the target common user and the items represented by the item representation data based on the fused user representation data and item representation data; and then, determining the items to be recommended for the target common user based on the interaction probability.

As an example, the above-mentioned execution entity can determine the interaction probability between the target shared user and each item in the item set; sort the items in descending order according to the interaction probability, so as to recommend a preset number of items in the front order as the target shared user's to-be-recommended items to the target shared user.

It should be noted that the various implementation methods in this embodiment can be implemented with reference to the corresponding implementation methods in the above-mentioned embodiment 200, and will not be described in detail here.

In this embodiment, a dual-target domain recommendation method is provided, in which bidirectional data migration is performed between the dual target domains, which solves the data sparsity problem in each target domain and achieves the purpose of benefiting enterprise users of the dual target domains from the data migration process; moreover, it is more in line with the actual situation under the dual target domain situation, and is conducive to the active participation of enterprise users of both parties corresponding to the dual target domains; in addition, based on the data migration process between the dual target domains, the accuracy of the recommendation results is improved.

Continuing to refer to FIG. 7 , a process 700 of another embodiment of a dual-target domain recommendation method is shown, including the following steps:

Step 701 : determining user representation data of the target common users and item representation data of the items according to interaction data between the target common users and the items in the item set corresponding to the dual target domains.

Step 702: transform the user representation data in the target domain into another target domain through an orthogonal mapping matrix to obtain the initial transformed user representation data corresponding to the target domain.

Step 703: Use local differential privacy technology to process the initial converted user representation data to obtain converted user representation data.

Step 704 , determining a gated selection vector according to the user representation data in the target domain and the converted user representation data corresponding to the other target domain through the trained embedding fusion module in the dual-target domain recommendation model.

Step 705 : performing weighted summation on the user representation data in the target domain and the converted user representation data corresponding to another target domain according to the gated selection vector to obtain fused user representation data.

Step 706 , using the trained dual-target domain recommendation model, according to the fused user representation data and item representation data, determine the items to be recommended for the target common user.

It can be seen from this embodiment that, compared with the embodiment corresponding to Figure 6, the process 700 of the dual-target domain recommendation method in this embodiment specifically illustrates the conversion process, hiding process, and fusion process of the user representation data. While ensuring the bidirectional migration of the dual target domains and the accuracy of the recommendation results, the security of the data in the dual target domains is further improved.

Continuing to refer to FIG8 , as an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a training device for a dual-target domain recommendation model. The device embodiment corresponds to the method embodiment shown in FIG2 , and the device can be specifically applied to various electronic devices.

As shown in FIG8 , the training device of the dual-target domain recommendation model performs a training operation for each target domain in the dual-target domains through the following units: a first determining unit 801 is configured to determine the user representation data of the common users and the item representation data of the items in the target domain according to the interaction data between the common users in the common user set corresponding to the dual-target domains and the items in the item set; a first obtaining unit 802 is configured to convert the user representation data in the target domain to another target domain, The converted user representation data corresponding to the target domain is obtained; the second obtaining unit 803 is configured to fuse the user representation data in the target domain representing the same shared user with the converted user representation data corresponding to another target domain to obtain the fused user representation data; the training unit 804 is configured to adopt a machine learning method, with the fused user representation data and the item representation data as input, and with the label representing whether the shared user corresponding to the input fused user representation data and the item corresponding to the input item representation data interact as the expected output, to train and obtain a dual-target domain recommendation model.

In some optional implementations of the present embodiment, the first obtaining unit 802 is further configured to: convert the user representation data in the target domain to another target domain through a trainable orthogonal mapping matrix to obtain the converted user representation data corresponding to the target domain.

In some optional implementations of the present embodiment, the above-mentioned device also includes: a matrix updating unit (not shown in the figure), which is configured to obtain, after each iterative operation in the training process of the dual-target domain recommendation model, an orthogonal mapping matrix corresponding to the next iterative operation of the dual-target domain according to the orthogonal mapping matrix in the target domain and the orthogonal mapping matrix in the other target domain.

In some optional implementations of the present embodiment, the first obtaining unit 802 is further configured to: convert the user representation data in the target domain to another target domain through a trainable orthogonal mapping matrix to obtain the initial converted user representation data corresponding to the target domain; and use local differential privacy technology to process the initial converted user representation data to obtain converted user representation data.

In some optional implementations of the present embodiment, the second obtaining unit 803 is further configured to: determine a gated selection vector according to the similarity between the user representation data in the target domain representing the same shared user and the converted user representation data corresponding to another target domain through a trainable embedding fusion module; and fuse the user representation data in the target domain representing the same shared user with the converted user representation data corresponding to another target domain according to the gated selection vector to obtain fused user representation data.

In some optional implementations of this embodiment, the second obtaining unit 803 is further configured to: perform weighted summation of the user representation data in the target domain and the converted user representation data corresponding to another target domain that represent the same common user according to the gated selection vector to obtain the fusion Post-user representation data.

In some optional implementations of the present embodiment, the training unit 804 is further configured to: predict the interaction probability between the common users corresponding to the fused user representation data and the items corresponding to the item representation data based on the fused user representation data and the item representation data; determine the loss between the interaction probability and the label based on the objective function, the interaction probability, the label and a preset normalization parameter; and train a dual-target domain recommendation model based on the loss.

In some optional implementations of this embodiment, the above-mentioned interaction data is an interaction matrix that characterizes the interaction between common users in the common user set and projects in the project set; and the above-mentioned first determination unit 801 is further configured to: determine the user representation data of the common users and the project representation data of the projects in the target domain based on the interaction matrix by matrix decomposition.

In the present embodiment, a training device for a dual-target domain recommendation model is provided, in which bidirectional data migration is performed between the dual target domains, which solves the data sparsity problem in each target domain and achieves the purpose that both corporate users of the dual target domains benefit from the data migration process; moreover, it is more in line with the actual situation under the dual target domain situation, which facilitates the active participation of corporate users of both parties corresponding to the dual target domains; in addition, based on the data migration process between the dual target domains, the accuracy of the recommendation results is improved.

Continuing to refer to FIG9 , as an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a dual-target domain recommendation device, which corresponds to the method embodiment shown in FIG6 , and can be specifically applied to various electronic devices.

As shown in FIG9 , the dual-target domain recommendation device performs a recommendation operation for each target domain in the dual-target domain through the following units: a second determining unit 901 is configured to determine the user representation data of the target common users and the item representation data of the items in the item set according to the interaction data between the target common users corresponding to the dual-target domains and the items in the item set; a third obtaining unit 902 is configured to convert the user representation data in the target domain to another target domain to obtain the converted user representation data corresponding to the target domain; a fourth obtaining unit 903 is configured to fuse the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain the fused user representation data; and a recommendation operation is performed on each target domain in the dual-target domain by the following units: The recommendation unit 904 is configured to determine the to-be-recommended items of the target common user according to the fused user representation data and item representation data through the trained dual-target domain recommendation model.

In some optional implementations of the present embodiment, the third obtaining unit 902 is further configured to: convert the user representation data in the target domain to another target domain through the orthogonal mapping matrix trained in the dual-target domain recommendation model, and obtain the converted user representation data corresponding to the target domain.

In some optional implementations of the present embodiment, the third obtaining unit 902 is further configured to: convert the user representation data in the target domain to another target domain through an orthogonal mapping matrix to obtain the initial converted user representation data corresponding to the target domain; and use local differential privacy technology to process the initial converted user representation data to obtain converted user representation data.

In some optional implementations of the present embodiment, the fourth obtaining unit 903 is further configured to: determine a gating selection vector according to the user representation data in the target domain and the converted user representation data corresponding to another target domain through an embedded fusion module trained in the dual-target domain recommendation model; and fuse the user representation data in the target domain with the converted user representation data corresponding to another target domain according to the gating selection vector to obtain fused user representation data.

In some optional implementations of this embodiment, the fourth obtaining unit 903 is further configured to: perform weighted summation of the user representation data in the target domain and the converted user representation data corresponding to another target domain according to the gated selection vector to obtain fused user representation data.

In some optional implementations of this embodiment, the recommendation unit 904 is further configured to: predict the interaction probability between the target common user and the items represented by the item representation data based on the fused user representation data and item representation data; and determine the items to be recommended for the target common user based on the interaction probability.

In some optional implementations of the present embodiment, the second determination unit 901 is further configured to: determine the user representation data of the target common users and the project representation data of the projects in the target domain based on the interaction matrix by means of matrix decomposition, wherein the interaction matrix represents the interaction data between the common users in the common user set and the projects in the project set corresponding to the dual target domains.

In the present embodiment, a dual-target domain recommendation device is provided, which performs bidirectional data migration between the dual target domains, solves the data sparsity problem in each target domain, and achieves the purpose that both enterprise users of the dual target domains benefit from the data migration process; moreover, it is more in line with the actual situation under the dual target domain situation, and facilitates the active participation of enterprise users of both parties corresponding to the dual target domains; in addition, based on the data migration process between the dual target domains, the accuracy of the recommendation results is improved.

Referring to FIG10, a schematic diagram of the structure of a computer system 1000 suitable for implementing the device (e.g., the devices 101, 102, 103, 105 shown in FIG1) of the embodiment of the present disclosure is shown below. The device shown in FIG10 is only an example and should not bring any limitation to the functions and scope of use of the embodiment of the present disclosure.

As shown in FIG10 , the computer system 1000 includes a processor (e.g., a CPU, a central processing unit) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage portion 1008 into a random access memory (RAM) 1003. Various programs and data required for the operation of the system 1000 are also stored in the RAM 1003. The processor 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input/output (I/O) interface 1005 is also connected to the bus 1004.

The following components are connected to the I/O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I/O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as needed, so that a computer program read therefrom is installed into the storage section 1008 as needed.

In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer readable medium, the computer program including a computer program for executing the process. In such an embodiment, the computer program can be downloaded and installed from a network through the communication part 1009, and/or installed from the removable medium 1011. When the computer program is executed by the processor 1001, the above functions defined in the method of the present disclosure are performed.

It should be noted that the computer-readable medium of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

Computer program code for performing operations of the present disclosure may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the client computer, partially on the client computer, as a stand-alone software package, or partially on the client computer. The client computer may be connected to the remote computer or the server. ...

The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the device, method and computer program product according to various embodiments of the present disclosure. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and/or flow chart, and the combination of the boxes in the block diagram and/or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

The units involved in the embodiments described in the present disclosure may be implemented by software or by hardware. The described units may also be arranged in a processor. For example, they may be described as: a processor including a first determination unit, a first obtaining unit, a second obtaining unit, and a training unit; for another example, they may be described as: a processor including a second determination unit, a third obtaining unit, a fourth obtaining unit, and a recommendation unit. Among them, the names of these units do not constitute a limitation on the units themselves under certain circumstances. For example, the first obtaining unit may also be described as "a unit that converts the user representation data in the target domain into another target domain and obtains the converted user representation data corresponding to the target domain".

As another aspect, the present disclosure further provides a computer-readable medium, which may be included in the device described in the above embodiment; or may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by the device, the computer device: for each of the dual target domains In the target domain, the following training operations are performed: according to the interaction data between the common users in the common user set and the items in the item set corresponding to the dual target domains, the user representation data of the common users and the item representation data of the items in the target domain are determined; the user representation data in the target domain is converted to another target domain to obtain the converted user representation data corresponding to the target domain; the user representation data in the target domain representing the same common user is fused with the converted user representation data corresponding to another target domain to obtain the fused user representation data; a machine learning method is adopted, with the fused user representation data and the item representation data as input, and the label representing whether there is interaction between the common users corresponding to the input fused user representation data and the items corresponding to the input item representation data as the expected output, to train a dual target domain recommendation model. The computer device can also be made to: perform the following recommendation operations for each target domain in the dual target domains: determine the user representation data of the target common users and the project representation data of the projects based on the interaction data between the target common users corresponding to the dual target domains and the projects in the project set; convert the user representation data in the target domain to another target domain to obtain the converted user representation data corresponding to the target domain; fuse the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain the fused user representation data; and determine the projects to be recommended for the target common users based on the fused user representation data through the trained dual target domain recommendation model.

The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present disclosure is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above invention concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present disclosure (but not limited to) by each other.

Claims (19)

A training method for a dual-target domain recommendation model, for each target domain in the dual-target domain, performing the following training operations: Determine user representation data of the common users and item representation data of the items in the target domain according to interaction data between the common users in the common user set and the items in the item set corresponding to the dual target domains; Converting the user representation data in the target domain to another target domain to obtain converted user representation data corresponding to the target domain; fusing the user representation data in the target domain representing the same common user with the converted user representation data corresponding to another target domain to obtain fused user representation data; and A machine learning method is adopted, with the fused user representation data and item representation data as input, and the label representing whether there is interaction between the common users corresponding to the input fused user representation data and the items corresponding to the input item representation data as the expected output, to train a dual-target domain recommendation model. The method according to claim 1, wherein converting the user representation data in the target domain to another target domain to obtain converted user representation data corresponding to the target domain comprises: The user representation data in the target domain is converted to another target domain through a trainable orthogonal mapping matrix to obtain converted user representation data corresponding to the target domain. The method according to claim 2, further comprising: After each iterative operation in the training process of the dual-target domain recommendation model, an orthogonal mapping matrix corresponding to the next iterative operation of the dual-target domain is obtained according to the orthogonal mapping matrix in the target domain and the orthogonal mapping matrix in another target domain. The method according to claim 2, wherein the user representation data in the target domain is converted to another target domain by a trainable orthogonal mapping matrix to obtain the target domain The corresponding converted user representation data includes: The user representation data in the target domain is converted to another target domain through a trainable orthogonal mapping matrix to obtain initial converted user representation data corresponding to the target domain; and The initial converted user representation data is processed using a local differential privacy technology to obtain the converted user representation data. The method according to claim 1, wherein fusing the user representation data in the target domain representing the same common user with the converted user representation data corresponding to another target domain to obtain the fused user representation data comprises: determining, by a trainable embedding fusion module, a gated selection vector according to a similarity between user representation data in the target domain and converted user representation data corresponding to another target domain representing the same common user; and According to the gated selection vector, the user representation data in the target domain representing the same common user is fused with the converted user representation data corresponding to another target domain to obtain the fused user representation data. The method according to claim 5, wherein the fusing, according to the gated selection vector, the user representation data in the target domain representing the same common user with the converted user representation data corresponding to another target domain to obtain the fused user representation data comprises: According to the gated selection vector, a weighted sum is performed on the user representation data in the target domain representing the same shared user and the converted user representation data corresponding to another target domain to obtain the fused user representation data. The method according to any one of claims 1 to 6, wherein the dual-target domain recommendation model is obtained by training using a machine learning method, taking the fused user representation data and the item representation data as input, and taking a label representing whether there is interaction between the common users corresponding to the input fused user representation data and the items corresponding to the input item representation data as the expected output, comprising: Predicting, based on the fused user representation data and the item representation data, an interaction probability between a common user corresponding to the fused user representation data and an item corresponding to the item representation data; Determining the loss between the interaction probability and the label according to the interaction probability, the label and a preset normalization parameter through an objective function; and The dual-target domain recommendation model is trained based on the loss. The method according to claim 1, wherein the interaction data is an interaction matrix characterizing the interaction between the common users in the common user set and the items in the item set; and The step of determining the user representation data of the common users and the project representation data of the projects in the target domain according to the interaction data between the common users in the common user set and the projects in the project set corresponding to the dual target domains includes: By adopting a matrix decomposition method, based on the interaction matrix, the user representation data of the common users and the project representation data of the projects in the target domain are determined. A dual-target domain recommendation method performs the following recommendation operations for each target domain in the dual-target domain: Determine user representation data of the target common users and item representation data of the items in the item set according to interaction data between the target common users corresponding to the dual target domains; Converting the user representation data in the target domain to another target domain to obtain converted user representation data corresponding to the target domain; fusing the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain fused user representation data; and The trained dual-target domain recommendation model is used to determine the to-be-recommended items of the target common user according to the fused user representation data and the item representation data. The method according to claim 9, wherein the user in the target domain The representation data is converted into another target domain to obtain the converted user representation data corresponding to the target domain, including: The user representation data in the target domain is converted to another target domain through the orthogonal mapping matrix trained in the dual-target domain recommendation model to obtain the converted user representation data corresponding to the target domain. The method according to claim 10, wherein the converting the user representation data in the target domain to another target domain through the orthogonal mapping matrix trained in the dual-target domain recommendation model to obtain the converted user representation data corresponding to the target domain comprises: Converting the user representation data in the target domain to another target domain through the orthogonal mapping matrix to obtain initial converted user representation data corresponding to the target domain; and The initial converted user representation data is processed using a local differential privacy technology to obtain the converted user representation data. The method according to claim 9, wherein fusing the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain fused user representation data comprises: Determining a gated selection vector according to the user representation data in the target domain and the converted user representation data corresponding to another target domain through the trained embedding fusion module in the dual-target domain recommendation model; and According to the gated selection vector, the user representation data in the target domain is fused with the converted user representation data corresponding to another target domain to obtain the fused user representation data. The method according to claim 12, wherein the fusing, according to the gated selection vector, the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain the fused user representation data comprises: According to the gated selection vector, the user representation data in the target domain is compared with another target domain. The converted user representation data corresponding to the domain are weighted summed to obtain the fused user representation data. The method according to claim 9, wherein the trained dual-target domain recommendation model determines the to-be-recommended items of the target common user according to the fused user representation data and the item representation data, comprising: Predicting the interaction probability between the target common user and the item represented by the item representation data according to the fused user representation data and the item representation data; and According to the interaction probability, the to-be-recommended items of the target common user are determined. The method according to claim 9, wherein determining the user representation data of the target common users and the item representation data of the items in the item set according to the interaction data between the target common users corresponding to the dual target domains and the items in the item set comprises: By adopting the matrix decomposition method, based on the interaction matrix, the user representation data of the target common users and the project representation data of the projects in the target domain are determined, wherein the interaction matrix represents the interaction data between the common users in the common user set corresponding to the dual target domains and the projects in the project set. A training device for a dual-target domain recommendation model performs a training operation for each target domain in the dual-target domain through the following units: A first determining unit is configured to determine user representation data of common users and item representation data of items in the target domain according to interaction data between common users in the common user set corresponding to the dual target domains and items in the item set; A first obtaining unit is configured to convert the user representation data in the target domain into another target domain to obtain converted user representation data corresponding to the target domain; A second obtaining unit is configured to fuse the user representation data in the target domain and representing the same common user with the converted user representation data corresponding to another target domain to obtain fused user representation data; and The training unit is configured to adopt a machine learning method, with the fused user representation data and the item representation data as input, and a label representing whether there is interaction between the common users corresponding to the input fused user representation data and the items corresponding to the input item representation data as the expected output, to train a dual-target domain recommendation model. A dual-target domain recommendation device performs a recommendation operation for each target domain in the dual-target domain through the following units: A second determining unit is configured to determine user representation data of the target common users and item representation data of the items according to interaction data between the target common users corresponding to the dual target domains and the items in the item set; A third obtaining unit is configured to convert the user representation data in the target domain into another target domain to obtain converted user representation data corresponding to the target domain; a fourth obtaining unit, configured to fuse the user representation data in the target domain with the converted user representation data corresponding to another target domain to obtain fused user representation data; and The recommendation unit is configured to determine the to-be-recommended items of the target common user according to the fused user representation data through the trained dual-target domain recommendation model. A computer readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the method according to any one of claims 1 to 15 is implemented. An electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 15.