WO2023121736A1

WO2023121736A1 - Content recommendation based on graph enhanced collaborative filtering

Info

Publication number: WO2023121736A1
Application number: PCT/US2022/044935
Authority: WO
Inventors: Linjun SHOU; Ming GONG; Daxin Jiang; Shuguang LIU
Original assignee: Microsoft Technology Licensing, Llc.
Priority date: 2021-12-23
Filing date: 2022-09-27
Publication date: 2023-06-29
Also published as: CN116340645A

Abstract

The present disclosure proposes a method, apparatus and computer program product for content recommendation based on graph-enhanced collaborative filtering. A candidate content item representation of a candidate content item may be generated. A set of historical content item representations corresponding to a set of historical content items of a target user may be generated. A set of general interest representations for all users may be generated based on a set of meta-interests, each meta-interest representing an interest element. A user interest representation of the target user may be generated based on the set of historical content item representations and the set of general interest representations. A click probability of the target user clicking the candidate content item may be predicted based on the candidate content item representation and the user interest representation.

Description

CONTENT RECOMMENDATION BASED ON GRAPH ENHANCED

COLLABORATIVE FILTERING

BACKGROUND

With the development of network technology and the growth of network information, recommendation systems are playing an increasingly important role in many online services. Based on different recommended content, there are different recommendation systems, e.g., a movie recommendation system, a book recommendation system, a music recommendation system, a product recommendation system, etc. These recommendation systems usually capture an interest of a user, and predict content that the user is interested in based on the interest of the user and recommend the content to the user.

SUMMARY

This Summary is provided to introduce a selection of concepts that are further described below in the Detailed Description. It is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. Embodiments of the present disclosure propose a method, apparatus and computer program product for content recommendation based on graph-enhanced collaborative filtering. A candidate content item representation of a candidate content item may be generated. A set of historical content item representations corresponding to a set of historical content items of a target user may be generated. A set of general interest representations for all users may be generated based on a set of meta-interests, each meta-interest representing an interest element. A user interest representation of the target user may be generated based on the set of historical content item representations and the set of general interest representations. A click probability of the target user clicking the candidate content item may be predicted based on the candidate content item representation and the user interest representation.

It should be noted that the above one or more aspects comprise the features hereinafter fully described and particularly pointed out in the claims. The following description and the drawings set forth in detail certain illustrative features of the one or more aspects. These features are only indicative of the various ways in which the principles of various aspects may be employed, and this disclosure is intended to include all such aspects and their equivalents.

BRIEF DESCRIPTION OF THE DRAWINGS

The disclosed aspects will hereinafter be described in connection with the appended drawings that are provided to illustrate and not to limit the disclosed aspects.

FIG.l illustrates an exemplary process for content recommendation based on graph-enhanced collaborative filtering according to an embodiment of the present disclosure. FIG.2 illustrates an exemplary process for training a collaborative filtering model according to an embodiment of the present disclosure.

FIG.3 illustrates an exemplary process of predicting a click probability through employing a content-based filtering model according to an embodiment of the present disclosure.

FIG.4 illustrates an exemplary process for training a content-based filtering model according to an embodiment of the present disclosure.

FIG.5 illustrates another exemplary process for training a collaborative filtering model according to an embodiment of the present disclosure.

FIG.6 illustrates an exemplary process for generating a cross-system contrastive prediction loss according to an embodiment of the present disclosure.

FIG.7 is a flowchart of an exemplary method for content recommendation based on graph- enhanced collaborative filtering according to an embodiment of the present disclosure.

FIG.8 illustrates an exemplary apparatus for content recommendation based on graph-enhanced collaborative filtering according to an embodiment of the present disclosure.

FIG.9 illustrates an exemplary apparatus for content recommendation based on graph-enhanced collaborative filtering according to an embodiment of the present disclosure.

DETAILED DESCRIPTION

The present disclosure will now be discussed with reference to several example implementations. It is to be understood that these implementations are discussed only for enabling those skilled in the art to better understand and thus implement the embodiments of the present disclosure, rather than suggesting any limitations on the scope of the present disclosure.

Collaborative Filtering (CF) is a widely used recommendation technology. The collaborative filtering technology aims to identify a similar user that is similar to a target user, determine an interest of the target user through an interest of the similar user, and recommend content to the target user based on the determined interest of the target user. Herein, a target user may refer to a user for whom the content recommendation is performed. The collaborative filtering technology described above may be implemented through a machine learning model. Herein, a machine learning model employing the content recommendation method based on the collaborative filtering may be referred to as a collaborative filtering model. A click probability of the target user clicking each candidate content item in a set of candidate content items may be predicted through the collaborative filtering model, thereby obtaining a set of click probabilities. Herein, a content item may refer to an individual item with specific content. For example, a movie, a book, a piece of music, etc. may be referred to as a content item. A content item to be recommended to the target user may be determined through ranking the set of click probabilities.

The collaborative filtering model may generate a candidate content item representation of a candidate content item and a user interest representation of a target user, and predict a click probability of the target user clicking the candidate content item based on the candidate content item representation and the user interest representation. Generally, historical content items previously browsed, clicked or viewed by the target user may indicate a user interest of the target user, and thus the user interest representation of the target user may be generated based on representations of the historical content items. When generating a representation of a content item such as a candidate content item and a historical content item, a knowledge graph corresponding to the content item may be considered. A knowledge graph may contain a large number of nodes representing a large number of entities and a large number of edges representing rich relations among the entities. In the knowledge graph, a neighbor node adjacent to a specific node may be considered as structural information of the specific node, e.g., attribute information, and an edge from the neighbor node to the specific node may be considered as a relation of the neighbor node with respect to the specific node. Therefore, when the collaborative filtering technology is employed for content recommendation, the candidate content item representation and the user interest representation may be enriched with the knowledge graph, so as to more accurately predict the click probability of the target user clicking the candidate content item. The content recommendation method employing collaborative filtering based on knowledge graph may be referred to as a content recommendation method based on graph-enhanced collaborative filtering. Embodiments of the present disclosure propose improved content recommendation based on graph-enhanced collaborative filtering. A candidate content item representation of a candidate content item and a user interest representation of a target user may be generated, and a click probability of the target user clicking the candidate content item may be predicted based on the generated candidate content item representation and the user interest representation. When generating the user interest representation, a set of historical content item representations corresponding to a set of historical content items of a target user and a set of general interest representations for all users may be generated, and the user interest representation may be generated based on the set of historical content item representations and the set of general interest representations. A general interest representation may correspond to a general interest. Herein, a general interest may refer to an interest for all users rather than a specific user. The set of general interest representations proposed by the embodiments of the present disclosure may be generated based on a set of meta-interests. Herein, a meta-interest may refer to an interest element in a set of interest elements used to constitute a general interest for all users. Each meta-interest may have a corresponding trainable embedding. Generating a general interest representation with a meta- interest may more broadly capture different interest points in a general interest for all users. The improved content recommendation method based on graph-enhanced collaborative filtering described above may be implemented through a collaborative filtering model according to the embodiments of the present disclosure.

In an aspect, the embodiments of the present disclosure propose to generate a candidate content item representation of a candidate content item based on a knowledge graph corresponding to the candidate content item. For example, a set of neighbor nodes adjacent to a candidate content item and a set of edges corresponding to the set of neighbor nodes may be identified from the knowledge graph, and a candidate content item representation may be generated based on a set of neighbor node representations corresponding to the set of neighbor nodes, a set of relation representations corresponding to the set of edges, importance of the set of neighbor nodes to the candidate content item, etc. The importance of a neighbor node to a candidate content item may be used to control information passed from the neighbor node to the candidate content item. The neighbor node that is more important to the candidate content item may pass more information to the candidate content item, and thus may influence the candidate content item representation to a greater extent. Similarly, for each historical content item in the set of historical content items, a historical content item representation of the historical content item may be generated based on a knowledge graph corresponding to the historical content item.

In another aspect, the embodiments of the present disclosure propose to employ Soft Distance Correlation to enhance a diversity of a set of general interest representations used to generate a user interest representation when training a collaborative filtering model. For example, dimensionality of a set of general interest representations may be reduced, to obtain a set of reduced-dimensionality general interest representations; and distance correlation constraints may be applied to the set of reduced-dimensionality general interest representations, to enhance a diversity of the set of reduced-dimensionality general interest representations. Applying distance correlation constraints to a set of general interest representations may separate individual general interest representations from each other in the vector space, thus ensuring that each general interest representation is as distinct as possible, so that different general interest representations may characterize different interests from different aspects. Further, reducing the dimensionality of the set of general interest representations firstly and applying the distance correlation constraints only to the set of reduced-dimensionality general interest representations may ensure a diversity of general interest representations at low dimension, while maintaining flexibility at high dimension. In this way, different points of interest of the user may be learned more comprehensively, so that a more accurate user interest representation may be generated.

In yet another aspect, the embodiments of the present disclosure propose to determine negative content item samples through Reciprocal Ratio Negative Sampling when constructing a training dataset for training a collaborative filtering model. Herein, a negative content item sample may refer to a content item sample that a user is not interested in. A negative content item sample may be sampled from a set of candidate content item samples based on a popularity of each content item sample in the set of candidate content item samples. For example, a content item with more user interactions may be considered as a popular content item. There may be a small probability for a popular content item to be sampled as a negative content item sample. Different from the existing way of determining a negative content item sample through random sampling, determining a negative content item sample through reciprocal ratio negative sampling may avoid determining a content item that a user has not encountered before but may be interested in as a negative content item sample of the user, thereby a quality of the negative content item sample may be improved.

In still another aspect, the embodiments of the present disclosure propose to train a collaborative filtering model with a Content-based Filtering (CBF) model. When generating a representation of a candidate content item and/or a historical content item, a content-based filtering model may utilize unstructured information of the candidate content item and/or the historical content item, such as a text description that embodies content of the candidate content item and/or the historical content item. Historical content item representations may be further used to generate a user interest representation of a target user. Training a collaborative filtering model with a content-based filtering model enables knowledge learned by the content-based filtering model, e.g., knowledge obtained from unstructured information of content items, to be transferred to the collaborative filtering model. The collaborative filtering model may be trained with the content-based filtering model through Cross-System Contrastive Learning. This approach provides a lightweight way to fuse the collaborative filtering model and the content-based filtering model. The collaborative filtering model trained with the content-based filtering model may consider a text description of a candidate content item and/or a historical content item when generating a candidate content item representation and/or a user interest representation, thereby a richer candidate content item representation and/or user interest representation may be generated. In addition, for some new content items, e.g., a newly released movie, a newly published book, etc., even if there is less user interaction for them or they have not been covered in previous training, these content items may also be recommended based on their text descriptions, which may effectively solve a cold start issue.

FIG.1 illustrates an exemplary process 100 for content recommendation based on graph-enhanced collaborative filtering according to an embodiment of the present disclosure. In the process 100, a click probability 112 of a target user u clicking a candidate content item 102 i (i E Sf) may be predicted through a collaborative filtering model 110.

The candidate content item 102 i may be a content item from a set of candidate content items that may be recommended to the target user it. The candidate content item 102 i may include, e.g., movie, book, music, video, product information, news, etc.

A candidate content item representation 122 of the candidate content item 102 i may be generated through a content item encoder 120. The content item encoder 120 may be, e.g., a Gated Path Graph Convolution Network including L convolution layers. Preferably, the content item encoder 120 may generate the candidate content item representation 122 based on a knowledge graph 106. The knowledge graph 106 may be denoted as Q = (V, £), where V is a set of nodes in the knowledge graph and £ is a set of edges among the set of nodes. The knowledge graph 106 may be a knowledge graph corresponding to the candidate content item 102 i. Taking the candidate content item 102 i being a movie as an example, the knowledge graph 106 may be a knowledge graph including a set of nodes whose entity type is movie and a set of edges among these nodes. The candidate content items 102 i may correspond to a node ( i E V) in the knowledge graph 106 . A set of neighbor nodes JVj = {Vj \(Vi, r_Lj, Vj) E T} adjacent to the node

and a set of edges {r_tj E JI} corresponding to the set of neighbor nodes may be identified from the knowledge graph 106. At a convolution layer I + 1 (0 < I < L) of the gated path graph convolution network, a representation of the node

at Z + 1 may be generated based on representations at a convolution layer I of a set of neighbor nodes JVj of the node v_t, a set of relation representations corresponding to a set of edges among the node and the set of neighbor nodes JVj, importance of the set of neighbor nodes JVj to the node v_t, as shown in the following equation:

where e_r is a relation representation of a relation

from the node

to the neighbor node Vj, Ytj is a gated function that controls information passed from the neighbor node Vj to the node v_t, which may be used to weight the information passed from the neighbor node Vj to the node v_t. e® may be an initial representation of the node

or the candidate content item i, which may be a randomly initialized representation or a representation obtained through a known knowledge graph embedding method. Similarly, a relation representation e_r.. may be a randomly initialized

^7 representation or a representation obtained through a known knowledge graph embedding method. The gated function y^ may reflect the importance of the neighbor node Vj to the node v . The gated function llowing equation:

Ytj = ^(e/e_r )

where < (■) is the sigmoid function used to limit the gated value between 0 and 1. A neighbor node that is more important to the node may pass more information to the node v_t, and thus may influence the representation of the node v_t, i.e., the representation of the candidate content item Z, to a greater extent. It should be appreciated that although in the equation (1), the candidate content item representation is generated based on three factors including, e.g., neighbor node representations, relation representations, and the importance of the neighbor nodes to the candidate content item, in some embodiments, it is also possible to generate the candidate content item representation based only on any one or two of these three factors.

To overcome a over-smoothing issue of graph convolution, a final representation of the node v_t, i.e., the candidate content item representation 122 e , may be obtained through aggregating the representations of nodes te layers, as shown in the following equation:

To generate a user interest representation 142 of the target user it, a set of historical content items 104 of the target user u may be obtained. The set of historical content items 104 may include a plurality of historical content items previously browsed, clicked or viewed by the target user it, e.g., a historical content item 104-1 to a historical content item 104-C, where C is the number of historical content items. A historical content item may include, e.g., movie, book, music, video, product information, news, etc. The set of historical content items 104 of the target user u may indicate user interests of the target user u. Taking the historical content item being a movie as an example, movies that the target user u has watched before may indicate which movies the user is interested in.

A set of historical content item representations 132 corresponding to the set of historical content items 104 may be generated through a set of content item encoders 130. The set of content item encoders 130 may include, e.g., a content item encoder 130-1 to a content item encoder 130-C. The set of historical content item representations 132 may include, e.g., a historical content item representation 132-1 to a historical content item representation 132-C. Each content item encoder in the set of content item encoders 130 may have a similar structure to the content item encoder 120, e.g., a content item encoder may be a gated path graph convolution network including L convolution layers. The historical content item representations 132-1 to the historical content item 132-C may be generated through a process similar to the process of generating the candidate content item representation 122. For example, for each historical content item in the set of historical content items 104, a historical content item representation of the historical content item may be generated based on a knowledge graph corresponding to the historical content item, e.g., the knowledge graph 106.

Subsequently, a user interest representation 142 of the target user u may be generated through a user encoder 140. For example, a set of general interest representations for all users may be generated based on a set of general interests 108 J³ for all users. The set of general interests J³ may contain interactions with content items from all users. The user interest representation 142 of the target user u may be generated based on the set of historical content item representations 132 and the set of general interest representations.

According to the embodiments of the present disclosure, each general interest p G J³ in the set of general interests J³ may consist of a set of meta-interests . Each meta-interest m G may represent an an interest element. For each general interest p, a general interest representation e_p of the general interest p may be generated through the set of meta-interests , as shown in the following equation:

where e_m is a trainable embedding of the meta-interest m, and {ft_pm\m G } is a linear weight derived from trainable weights {P_pm\m G } for the general interest p, as shown in the following equation:

The operations described above may be performed for each general interest p G J³, to generate a set of general interest representations {e_p}. Next, a representation of the user interest at the convolution layer I may be generated through aggregating the set of historical content item representations 132 at the convolution layer I with the set of general interest representations {e_p}, as shown in the following equation:

where (u, i) G J³⁺ is a historical interaction of the user it, e® is a representation of the historical content item i at the convolution layer Z, and a_p is a weight for the general interest obtained through the attention mechanism, as shown in the following equation:

Similar to generating the candidate content item representation 122, to overcome the oversmoothing issue of graph convolution, a final representation of user interests, i.e. the user interest representation 142

may be obtained through aggregating the representations of user interests at all intermediate layers, as shown in the following equation:

After the candidate content item representation 122 e and the user interest representation 142 are generated, the click probability 112 of the target user it clicking the candidate content item 102 i may be predicted through a predicting layer 150. The click probability may be denoted as y . In an implementation, the click probability may be predicted through applying a dot product operation to the user interest representation 142 and the candidate content item representation 122 e , as shown in the following equation: yu^Ci = (^)^T ■ e (9)

Since the collaborative filtering model 110 may utilize a knowledge graph to generate the candidate content item representation and/or the historical content item representation, and may generate the user interest representation based on the set of meta-interests, the collaborative filtering model 110 may also be referred to as a Knowledge-Graph-Enhanced Meta-Interest Network.

It should be appreciated that the process for content recommendation based on graph-enhanced collaborative filtering described above in conjunction with FIG.1 is merely exemplary. Depending on actual application requirements, the steps in the process for content recommendation based on graph-enhanced collaborative filtering may be replaced or modified in any manner, and the process may include more or fewer steps. For example, although in the process 100, the knowledge graph corresponding to the candidate content item and/or the historical content item is considered when generating the candidate content item representation and/or the historical content item representation, in some embodiments, it is also feasible not to consider the knowledge graph. In this case, the corresponding candidate content item representation and/or historical content item representation may be generated based only on the candidate content item and/or the historical content item itself. In addition, the specific order or hierarchy of the steps in the process 100 is merely exemplary, and the process for content recommendation based on graph-enhanced collaborative filtering may be performed in an order different from the described one.

A collaborative filtering model, e.g., the collaborative filtering model 110 in FIG.1, may be trained through a number of approaches. FIG.2 illustrates an exemplary process 200 for training a collaborative filtering model according to an embodiment of the present disclosure. When a collaborative filtering model trained through the process 200 is actually deployed, a click probability of a target user clicking a candidate content item may be predicted.

As described in conjunction with FIG. l, a set of general interest representations for all users is considered when generating a user interest representation of the target user. The embodiments of the present disclosure propose to employ soft distance correlation to enhance a diversity of the set of general interest representations.

For example, at 202, dimensionality of the set of general interest representations may be reduced, to obtain a set of reduced-dimensionality general interest representations. In an implementation, the dimensionality of the set of general interest representations may be reduce with Principal Component Analysis (PCA), as shown in the following equation:

where e is a ratio of principal component to be maintained after PCA. The ratio may be a value between 0 and 1.

At 204, a soft distance correlation prediction loss - softDCorr may be generated through applying distance correlation constraints to the set of reduced-dimensionality general interest representations, as shown in the following equation:

where DCov ) is used to calculate distance covariance, and DVar ) is used to calculate distance variance. It should be appreciated that when in Equation (10), e = 1, £_SoftDCorr produces an original distance correlation prediction loss.

Applying the distance correlation constraints to the set of general interest representations may separate individual general interest representations from each other in the vector space, thus ensuring that each general interest representation is as distinct as possible, so that different general interest representations may characterize different interests from different aspects. Further, reducing the dimensionality of the set of general interest representations firstly and applying the distance correlation constraints only to the set of reduced-dimensionality general interest representations may ensure a diversity of general interest representations at low dimension, while maintaining flexibility at high dimension. In this way, different points of interest of the user may be learned more comprehensively, so that a more accurate user interest representation may be generated.

Additionally or alternatively, the collaborative filtering model may be trained with a training dataset including a plurality of positive content item samples and a plurality of negative content item samples.

For example, at 210, a plurality of positive content item samples {i ⁺] may be obtained. For example, content items previously browsed, clicked or viewed by a user it in a set of candidate content item samples may be considered as positive content item samples i⁺ for the user u.

At 212, a plurality of negative content item samples {i “ } may be obtained. The embodiments of the present disclosure propose to determine a negative content item sample i~ through reciprocal ratio negative sampling. For example, a plurality of negative content item samples {i “ } may be sampled from a set of candidate content item samples based on a popularity of each sample in the set of candidate content item samples. For example, a content item with more user interactions may be considered as a popular content item. There may be a small probability for a popular content item to be sampled as a negative content item sample i~ . For a content item sample i E S a probability of it being sampled as a negative content item sample i~ may be given by the following equation: ₍₁₂₎ where c(i) represents a count of interactions with the content item sample i from all users. Different from the existing way of determining a negative content item sample through random sampling, determining a negative content item sample through reciprocal ratio negative sampling may avoid determining a content item that a user has not encountered before but may be interested in as a negative content item sample of the user, thereby a quality of the negative content item sample may be improved.

At 214, a training dataset 'll = {(it, i⁺, i^_)|(it, i⁺) G J’⁺, (u, i~)

} may be constructed based on the plurality of positive content item samples {i ⁺] and the plurality of negative content item samples {i “ }.

At 216, pairwise Bayesian Personalized Ranking (BPR) prediction losses £_BPR may be generated based on the constructed dataset. The BPR prediction loss aims to employ a concept of contrastive learning to assign a higher score to a content item viewed by a user than a content item that the user is not interested in. The BPR prediction loss £_BPR may be generated, e.g., by the following equation:

Alternatively or additionally, at 218, a L2 regularization prediction loss || 0|| ₂ may be generated based on the constructed dataset, where 0 =

(u, i⁺, i~) G 'll}, and II0II2 is the L2- norm of the user interest representation/content item representation.

Subsequently, at 220, a comprehensive prediction loss £_CF for the collaborative filtering model may be generated based on at least one of the distance correlation prediction loss, the BPR prediction loss, and the L2 regularization prediction loss, as shown in the following equation:

and A₂ are hyper-parameters used to control the prediction loss weights.

At 222, the collaborative filtering model may be trained through minimizing the comprehensive prediction loss L_CP.

It should be appreciated that the process for training the collaborative filtering model described above in conjunction with FIG.2 is merely exemplary. Depending on actual application requirements, the steps in the process for training the collaborative filtering model may be replaced or modified in any manner, and the process may include more or fewer steps. For example, although in the process 200, the distance correlation prediction loss, the BPR prediction loss, and the L2 regularization prediction loss are considered when generating the comprehensive prediction loss of the collaborative filtering model, in some embodiments, the comprehensive prediction loss of the collaborative filtering model may be generated based only on any one or two of the distance correlation prediction loss, the BPR prediction loss, and the L2 regularization prediction loss. In addition, the specific order or hierarchy of the steps in the process 200 is merely exemplary, and the process for training the collaborative filtering model may be performed in an order different from the described one.

It is described above in conjunction with FIG. l that a candidate content item representation of a candidate content item and a user interest representation of a target user are generated through a collaborative filtering model, and further a click probability of the target user clicking the candidate content item is predicted based on the generated candidate content item representation and interest representation. When generating the candidate content item representation and/or the user interest representation, the collaborative filtering model considers structured information of a content item, e.g., various attribute information from a knowledge graph. The embodiments of the present disclosure propose to train the collaborative filtering model with a content-based filtering model. Similar to the collaborative filtering model, the content-based filtering model may generate a candidate content item representation of a candidate content item and a user interest representation of a target user, and predict a click probability of the target user clicking the candidate content item based on the candidate content item representation and the user interest representation. Furthermore, the user interest representation may also be generated based on historical content item representations of the user. However, when generating a representation of a candidate content item and/or a historical content item, the content-based filtering model may utilize unstructured information of the candidate content item and/or the historical content item, such as a text description that embodies content of the candidate content item and/or the historical content item. The collaborative filtering model may be trained with the content-based filtering model through cross-system contrastive learning, which provides a lightweight way to fuse the collaborative filtering model and the content-based filtering model. The collaborative filtering model trained with the content-based filtering model may consider a text description of a candidate content item and/or a historical content item when generating a candidate content item representation and/or a user interest representation, thereby a richer candidate content item representation and/or user interest representation may be generated. In addition, for some new content items, e.g., a newly released movie, a newly published book, etc., even if there is less user interaction for them or they have not been covered in previous training, these content items may also be recommended based on their text descriptions, which may effectively solve a cold start issue.

FIG.3 illustrates an exemplary process 300 of predicting a click probability through employing a content-based filtering model according to an embodiment of the present disclosure. In the process 300, a click probability 312 of a target user it clicking a candidate content item 302 i may be predicted through a content-based filtering model 310.

The candidate content item 302 i may be a content item from a set of candidate content items that may be recommended to the target user it. The candidate content item 302 i may include, e.g., movie, book, music, video, product information, news, etc. A text description 306 x_t of the candidate content item 302 i may be obtained. The text description 306 x_t may be obtained in a number of known ways. In an implementation, the text description 306 x_t may be extracted from a knowledge graph corresponding to the candidate content item 302 i. In another implementation, articles related to the candidate content item 302 i may be searched from the Internet, and important paragraphs may be identified from the searched articles as the text description 306 x_t of the candidate content item 302 i through a trained machine learning model.

A candidate content item representation 322

of the candidate content item 302 i may be generated based on the text description 306 x_t through a content item encoder 320. The content item encoder 320 may be, e.g., a pre-trained transformer-based model, such as a Bidirectional Encoder Representations from Transformers (BERT) model, a Generative Pre-trained Transformer 2 (GPT-2) model, etc. Taking the content item encoder 320 being a BERT model as an example, the candidate content item representation 322

may be represented by the following equation: e_£ = B£7?T(x_£) e R^ft (15)

A set of historical content items 304 of the target user u may be obtained. The set of historical content items 304 may include a plurality of historical content items previously browsed, clicked or viewed by the target user it, e.g., a historical content item 304-1 to a historical content item 304-B, where B is the number of historical content items. A historical content item may include, e.g., movie, book, music, video, product information, news, etc. The set of historical content items 304 of the target user u may indicate user interests of the target user u. A set of text descriptions 308 corresponding to the set of historical content items 304 may be obtained. For example, a text description 308-1 of the historical content item 304-1 may be obtained, a text description 308-B of the historical content item 304-B may be obtained, etc.

A set of historical content item representations 332 corresponding to the set of historical content items 304 may be generated through a set of content item encoders 330. The set of content item encoders 330 may include, e.g., a content item encoder 330-1 to a content item encoder 330-B. The set of historical content item representations 332 may include, e.g., a historical content item representation 332-1 to a historical content item 332-B . Each content item encoder in the set of content item encoders 330 may have a similar structure to the content item encoder 320, e.g., the content item encoder may be a pre-trained transformer-based model such as a BERT model, a GPT-2 model, etc. The historical content item representations 332-1 to the historical content item 332-B may be generated through a process similar to the process of generating the candidate content item representation 322. For example, for each historical content item in the set of historical content items 304, a historical content item representation of the historical content item may be generated based on a text description of the historical content item.

Subsequently, a user interest representation 342 e_u of the target user u may be generated through a user encoder 340. The user encoder 340 may generate the user interest representation 342 e_u based on the set of historical content item representations 332 of the target user it . In an implementation, the user interest representation 342 e_u may be generated through a weighted sum of B historical content item representations included in the set of historical content item representations 332, as shown in the following equation:

where a_b is a attention weight assigned to a historical content item i_{u b}, which is obtained by passing features through two linear layers, as shown in the following equation:

A = tanh(E_uA_/Ci + b_/Ci) A_fC2 + b_fCz G R^Bxl (18) where

G R¹ are weights and biases of two fully- connected layers, respectively.

After the candidate content item representation 322 and the user interest representation 342 e_u are generated, the click probability 312 of the target user it clicking the candidate content item 302 may be predicted through a predicting layer 350. The click probability may be denoted as aⁿ implementation, the click probability y^{F F} may be predicted through applying a dot product operation to the user interest representation 342 e_u and the candidate content item representation 322 equation: yu^Ci ^F = (e_u)^T ■ e_t

Since the content-based filtering model 310 employs a multi-head self-attention mechanism, it may also be referred to as a Neural Recommendation with Multi-Head Self-Attention (NRMS) model. In particular, in the case where the content-based filtering model 310 employs a pre-trained BERT model as a content item encoder, the content-based filtering model 310 may also be referred to as an NRMS-BERT model.

It should be appreciated that the process for predicting the click probability through employing the content-based filtering model described above in conjunction with FIG.3 is merely exemplary. Depending on actual application requirements, the steps in the process for predicting the click probability through employing the content-based filtering model may be replaced or modified in any manner, and the process may include more or fewer steps. In addition, the specific order or hierarchy of the steps in the process 300 is merely exemplary, and the process for predicting the click probability through employing the content-based filtering model may be performed in an order different from the described one.

According to the embodiments of the present disclosure, a collaborative filtering model may be trained with a content-based filtering model. For example, the collaborative filtering model 110 in FIG.1 may be trained with the content-based filtering model 310 in FIG.3. Preferably, the content-based filtering model may be pre-trained before the collaborative filtering model is trained with the content-based filtering model. A negative sampling method may be employed to train the content-based filtering model. FIG.4 illustrates an exemplary process 400 for training a contentbased filtering model according to an embodiment of the present disclosure. When a content-based filtering model trained through the process 400 is actually deployed, a click probability of a target user clicking a candidate content item may be predicted.

Firstly, a training dataset for training a content-based filtering model may be constructed. In an implementation, the training dataset may be constructed through employing a list-wise strategy. For example, at 410, a plurality of positive content item samples may be obtained. For example, content items previously browsed, clicked or viewed by a user in a set of candidate content item samples may be considered as positive content item samples.

At 420, a plurality of negative content item sample sets corresponding to the plurality of positive content item samples may be obtained. For example, for each positive content item sample, a content item sample set that is presented in the same session as the positive content item sample but has not been clicked by the user may be regarded as a negative content item sample set corresponding to the positive content item sample.

At 430, a training dataset for training a content-based filtering model may be constructed based on the plurality of positive content item samples and the plurality of negative content item sample sets corresponding to the plurality of positive content item samples.

Subsequently, a plurality of posterior click probabilities corresponding to the plurality of positive content item samples may be generated. For example, at 440, a positive content item sample click probability corresponding to each positive content item sample may be predicted. A positive content item sample click probability corresponding to the i-th positive content item sample may be denoted as y .

At 450, for each negative content item sample in a negative content item sample set corresponding to the positive content item sample, a negative content item sample click probability corresponding to the negative content item sample may be predicted, to obtain a negative content item sample click probability set corresponding to the negative content item sample set. A negative content item sample click probability set corresponding to the negative content item sample set of the i-th positive content item sample may be denoted as [y^, y^₂> — > yijc], where K is the number of negative content item samples included in the negative content item sample click probability set. In this way, the click probability predicting problem may be expressed as a pseudo K+l way classification task.

At 460, a posterior click probability corresponding to the positive content item sample may be calculated based on the positive content item sample click probability and the negative content item sample click probability set. A posterior click probability corresponding to the i-th positive content item sample may be denoted as pt. In an implementation, the posterior click probability corresponding to the positive content item sample may be calculated through normalizing the positive content item sample click probability y* and the negative content item sample click probability set [y i. y^' ■■■ > 9^] using a softmax function, as shown in the following equation:

The operations from the step 440 to the step 460 described above may be performed for each of the plurality of positive content item samples in the training dataset, so that at 470, a plurality of posterior click probabilities corresponding to the plurality of positive content item samples may be obtained.

At 480, a prediction loss may be generated based on the plurality of posterior click probabilities. In an implementation, the prediction loss may be generated through calculating a negative loglikelihood of the plurality of posterior click probabilities, as shown in the following equation:

is a positive content item sample set composed of the plurality of positive content item samples.

At 490, a content-based filtering model may be optimized through minimizing the prediction loss. It should be appreciated that the process for training the content-based filtering model described above in conjunction with FIG.4 is merely exemplary. Depending on actual application requirements, the steps in the process for training the content-based filtering model may be replaced or modified in any manner, and the process may include more or fewer steps. For example, although in the process 400, the list-wise strategy is employed to construct the training dataset, other approaches may be employed to construct the training dataset, e.g., the Reciprocal Ratio Negative Sampling method described in conjunction with FIG.2 may also be employed when determining negative content item samples. In addition, the specific order or hierarchy of the steps in the process 400 is merely exemplary, and the process for training the content-based filtering model may be performed in an order different from the described one. The embodiments of the present disclosure propose to train a collaborative filtering model with a content-based filtering model through cross-system contrastive learning. FIG.5 illustrates another exemplary process 500 fortraining a collaborative filtering model according to an embodiment of the present disclosure. Compared to the process 200 described in conjunction with FIG.2, in the process 500, in addition to generating a comprehensive prediction loss of a collaborative filtering model based on at least one of a soft distance correlation prediction loss, a BPR prediction loss, and a L2 regularization prediction loss, a cross-system contrastive prediction loss may also be generated through the collaborative filtering model and a content-based filtering model employing cross-system contrastive learning.

A step 502 to a step 520 may correspond to the step 202 to the step 220 in FIG.2, respectively.

At 530, a cross-system contrastive prediction loss £_cs may be generated through a collaborative filtering model and a content-based filtering model, employing cross-system contrastive learning. An exemplary process for generating the cross-system contrastive prediction loss £_cs will be described later in conjunction with FIG.6.

At 540, a cross-system enhanced comprehensive prediction loss £_CCF of the collaborative filtering model may be generated based on the comprehensive prediction loss £_CF and the cross-system contrastive prediction loss £_cs, as shown in the following equation: CCF ⁼ ^CF + ^cs- es (22) where _cs is a weight used to control the cross-system contrastive prediction loss £_cs.

At 550, the collaborative filtering model may be optimized through minimizing the cross-system enhanced comprehensive prediction loss £_CCF-

It should be appreciated that the process for training the collaborative filtering model with the content-based filtering model described above in conjunction with FIG.5 is merely exemplary. Depending on actual application requirements, the steps in the process for training the collaborative filtering model with the content-based filtering model may be replaced or modified in any manner, and the process may include more or fewer steps. For example, while in process 500 the collaborative filtering model is trained based on both the comprehensive prediction loss and the cross-system contrastive prediction loss, in some embodiments, the collaborative filtering model may be trained based only on the cross-system contrastive prediction loss. In this case, the collaborative filtering model may be optimized through minimizing the cross-system contrastive prediction loss. In addition, the specific order or hierarchy of the steps in the process 500 is merely exemplary, and the process for training the collaborative filtering model with the content-based filtering model may be performed in an order different from the described one.

FIG.6 illustrates an exemplary process 600 for generating a cross-system contrastive prediction loss according to an embodiment of the present disclosure. The process 600 may correspond to the step 530 in FIG.5.

A cross-system contrastive prediction loss 660 £_cs may be generated with a pre-constructed training dataset 610. The training dataset 610 may be, e.g., a training dataset U = {(it, i⁺, i^_)|(it, i⁺) E J^>+, (it, i-)

} constructed through the step 510 to the step 514 in FIG.5, i.e., the step 210 to the step 214 in FIG.2. For each training data 620 d = (it, i⁺, i~) in the training dataset 610 Ft, a cross-system contrastive sub -predict! on loss 650 L^s f°^r that training data 620 d may be generated.

The training data 620 d may include, e.g., a user sample 622 it, a positive content item sample 624 t⁺, and a negative content item sample 626 i~ . The positive content item sample 624 i⁺ and negative content item sample 626 i~ may be associated with the user sample 622 u. For example, the positive content item sample 624 i⁺ may be a content item viewed by the user sample 622 it, and the negative content item sample 626 i~ may be a content item that the user sample 622 u is not interested in.

A collaborative filtering model 630 may correspond to, e.g., the collaborative filtering model 110 in FIG. l. A first user sample interest representation 632 e ^F of the user sample 622 it, a first positive content item sample representation 634 e^FF of the positive content item sample 624 i⁺, and a first negative content item sample representation 636 ef- of the negative content item sample 626 i~ may be generated through the collaborative filtering model 630.

A content-based filtering model 640 may correspond to, e.g., the content-based filtering model 310 in FIG.3. A second user sample interest representation 642 e ^BF of the user sample 622 it, a second positive content item sample representation 644 e^F+ ^F of the positive content item sample 624 t⁺, and a second negative content item sample representation 646 e^F- ^F of the negative content item sample 626 i~ may be generated through the content-based filtering model 640. The contentbased filtering model 640 may generate the second positive content item sample representation 644 e^F+ ^F based on a text description of the positive content item sample 624 i ⁺ . Similarly, the content-based filtering model 640 may generate the second negative content item sample representation 646 e^F- ^F based on a text description of the negative content item sample 626 i~ . Subsequently, a cross-system contrastive sub -prediction loss 650 LQ_S for the training data 620 d may be generated based on the first user sample interest representation 632 e ^F, the first positive content item sample representation 634 e^FF , the first negative content item sample representation 636 e^F- the second user sample interest representation 642 e^^F , the second positive content item sample representation 644 e^F+ ^F, and the second negative content item sample representation 646 e^F- ^F . For example, a cross-system contrastive learning may be employed to generate the cross- system contrastive sub -prediction loss 650 £^s, ^as shown in the following equation: eg))) (23)

The operations described above may be performed for each of the plurality of training data in the training dataset 1Z, to obtain a plurality of cross-system contrastive sub -prediction losses £Q_S- The plurality of cross-system contrastive sub -prediction losses £Q_S may be combined into a crosssystem contrastive prediction loss 660 £_cs of the collaborative filtering model 630, as shown in the following equation:

The cross-system contrastive prediction loss £_cs may guide the collaborative filtering model to fuse content-sensitive information from the content-based filtering model, e.g., information related to text descriptions. For example, the cross-system contrastive prediction loss £_cs may guide the collaborative filtering model to generate a content item representation that interacts not only with a user interest representation output by the collaborative filtering model itself, but also with a user interest representation output by the content-based filtering model. Similarly, the crosssystem contrastive prediction loss £_cs may guide the collaborative filtering model to generate a user interest representation that interacts not only with a content item representation output by the collaborative filtering model itself, but also with a content item representation output by the content-based filtering model. This approach may guide the collaborative filtering model to learn a representation output by a content-based filtering model that considers the text description of the content item when generating a content item representation and/or a user interest representation. Therefore, the knowledge learned by the content-based filtering model may be transferred to the collaborative filtering model.

It should be appreciated that the process for generating the cross-system contrastive prediction loss described above in conjunction with FIG.6 is merely exemplary. Depending on actual application requirements, the steps in the process for generating the cross-system contrastive prediction loss may be replaced or modified in any manner, and the process may include more or fewer steps. In addition, the specific order or hierarchy of the steps in the process 600 is merely exemplary, and the process for generating the cross-system contrastive prediction loss may be performed in an order different from the described one.

The content recommendation based on graph-enhanced collaborative filtering according to the embodiments of the present disclosure may be applicable to various types of recommendations, e.g., movie recommendation, book recommendation, music recommendation, video recommendation, product recommendation, news recommendation, etc. In particular, when a collaborative filtering model according to the embodiments of the present disclosure, such as the collaborative filtering model 110 shown in FIG.l, is used for online deployment, since the collaborative filtering model fully considers user interactions, it is especially suitable for recommendations in relatively static fields with slow content updates or with many user interactions, e.g., movie recommendation, book recommendation, music recommendation, video recommendation, product recommendation, etc. When a content-based filtering model according to the embodiments of the present disclosure, such as the content-based filtering model 310 shown in FIG.3, is used for online deployment, since the content-based filtering model may make recommendations based on text information of content items, it is especially suitable for recommendation in relatively dynamic fields with fast content updates or with less user interactions, e.g., news recommendation, etc.

FIG.7 is a flowchart of an exemplary method 700 for content recommendation based on graph- enhanced collaborative filtering according to an embodiment of the present disclosure.

At 710, a candidate content item representation of a candidate content item may be generated.

At 720, a set of historical content item representations corresponding to a set of historical content items of a target user may be generated.

At 730, a set of general interest representations for all users may be generated based on a set of meta-interests, each meta-interest representing an interest element.

At 740, a user interest representation of the target user may be generated based on the set of historical content item representations and the set of general interest representations.

At 750, a click probability of the target user clicking the candidate content item may be predicted based on the candidate content item representation and the user interest representation.

In an implementation, the generating a candidate content item representation may comprise: generating the candidate content item representation based on a knowledge graph corresponding to the candidate content item.

The generating the candidate content item representation may comprise: identifying a set of neighbor nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighbor nodes from the knowledge graph; and generating the candidate content item representation based on at least one of: a set of neighbor node representations corresponding to the set of neighbor nodes, a set of relation representations corresponding to the set of edges, and importance of the set of neighbor nodes to the candidate content item.

In an implementation, the generating a set of historical content item representations may comprise, for each historical content item in the set of historical content items: generating a historical content item representation of the historical content item based on a knowledge graph corresponding to the historical content item.

In an implementation, the generating a user interest representation may comprise: generating the user interest representation through aggregating the set of historical content item representations with the set of general interest representations.

In an implementation, the method 700 may be performed through a collaborative filtering model. A training of the collaborative filtering model may comprise: reducing dimensionality of the set of general interest representations, to obtain a set of reduced-dimensionality general interest representations; and applying distance correlation constraints to the set of reduced-dimensionality general interest representations, to enhance a diversity of the set of reduced-dimensionality general interest representations.

The reducing dimensionality of the set of general interest representations may comprise: reducing the dimensionality of the set of general interest representations with Principal Component Analysis.

A training of the collaborative filtering model may comprise: training the collaborative filtering model with a training dataset including a plurality of positive content item samples and a plurality of negative content item samples, wherein the plurality of negative content item samples are sampled from a set of candidate content item samples based on a popularity of each sample in the set of candidate content item samples.

A training of the collaborative filtering model may comprise: training the collaborative filtering model with a content-based filtering model.

The training the collaborative filtering model with a content-based filtering model may comprise: generating, through the collaborative filtering model and the content-based filtering model, a cross-system contrastive prediction loss employing cross-system contrastive learning; and optimizing the collaborative filtering model at least through minimizing the cross-system contrastive prediction loss.

The generating a cross-system contrastive prediction loss may comprise: for each training data of multiple training data included in a training dataset, generating a cross-system contrastive sub prediction loss for the training data, to obtain multiple cross-system contrastive sub prediction losses for the multiple training data; and combining the multiple cross-system contrastive sub prediction losses into the cross-system contrastive prediction loss.

The training data may include: a user sample, and a positive content item sample and a negative content item sample associated with the user sample. The generating a cross-system contrastive sub prediction loss may comprise: generating, through the collaborative filtering model, a first user sample interest representation of the user sample, a first positive content item sample representation of the positive content item sample, and a first negative content item sample representation of the negative content item sample; generating, through the content-based filtering model, a second user sample interest representation of the user sample, a second positive content item sample representation of the positive content item sample, and a second negative content item sample of the negative content item sample; and generating the cross-system contrastive sub prediction loss based on the first user sample interest representation, the first positive content item sample representation, the first negative content item sample representation, the second user sample interest representation, the second positive content item sample representation and the second negative content item sample representation.

The generating a second positive content item sample representation may comprise: generating, through the content-based filtering model, the second positive content item sample representation based on a text description of the positive content item sample. The generating a second negative content item sample representation may comprise: generating, through the content-based filtering model, the second negative content item sample representation based on a text description of the negative content item sample.

The candidate content item or the historical content item may include at least one of movie, book, music, video, product information, and news.

It should be appreciated that the method 700 may further include any steps/processes for content recommendation based on graph-enhanced collaborative filtering according to the embodiments of the present disclosure as mentioned above.

FIG.8 illustrates an exemplary apparatus 800 for content recommendation based on graph- enhanced collaborative filtering according to an embodiment of the present disclosure.

The apparatus 800 may comprise: a candidate content item representation generating module 810, for generating a candidate content item representation of a candidate content item; a historical content item representation generating module 820, for generating a set of historical content item representations corresponding to a set of historical content items of a target user; a general interest representation generating module 830, for generating a set of general interest representations for all users based on a set of meta-interests, each meta-interest representing an interest element; a user interest representation generating module 840, for generating a user interest representation of the target user based on the set of historical content item representations and the set of general interest representations; and a click probability predicting module 850, for predicting a click probability of the target user clicking the candidate content item based on the candidate content item representation and the user interest representation. Moreover, the apparatus 800 may further comprise any other modules configured for content recommendation based on graph-enhanced collaborative filtering according to the embodiments of the present disclosure as mentioned above. FIG.9 illustrates an exemplary apparatus 900 for content recommendation based on graph- enhanced collaborative filtering according to an embodiment of the present disclosure.

The apparatus 900 may comprise at least one processor 910 and a memory 920 storing computer- executable instructions. The computer-executable instructions, when executed, may cause the at least one processor 910 to: generate a candidate content item representation of a candidate content item; generate a set of historical content item representations corresponding to a set of historical content items of a target user; generate a set of general interest representations for all users based on a set of meta-interests, each meta-interest representing an interest element; generate a user interest representation of the target user based on the set of historical content item representations and the set of general interest representations; and predict a click probability of the target user clicking the candidate content item based on the candidate content item representation and the user interest representation.

It should be appreciated that the processor 910 may further perform any other steps/processes of the method for content recommendation based on graph-enhanced collaborative filtering according to the embodiments of the present disclosure as mentioned above.

The embodiments of the present disclosure propose a computer program product for content recommendation based on graph-enhanced collaborative filtering, comprising a computer program that is executed by at least one processor for: generating a candidate content item representation of a candidate content item; generating a set of historical content item representations corresponding to a set of historical content items of a target user; generating a set of general interest representations for all users based on a set of meta-interests, each meta-interest representing an interest element; generating a user interest representation of the target user based on the set of historical content item representations and the set of general interest representations; and predicting a click probability of the target user clicking the candidate content item based on the candidate content item representation and the user interest representation. In addition, the computer program may further be performed for implementing any other steps/processes of the method for content recommendation based on graph-enhanced collaborative filtering according to the embodiments of the present disclosure as mentioned above.

The embodiments of the present disclosure may be embodied in non-transitory computer-readable medium. The non-transitory computer readable medium may comprise instructions that, when executed, cause one or more processors to perform any operation of the method for content recommendation based on graph-enhanced collaborative filtering according to the embodiments of the present disclosure as mentioned above.

It should be appreciated that all the operations in the methods described above are merely exemplary, and the present disclosure is not limited to any operations in the methods or sequence orders of these operations, and should cover all other equivalents under the same or similar concepts. In addition, the articles “a” and “an” as used in this specification and the appended claims should generally be construed to mean “one” or “one or more” unless specified otherwise or clear from the context to be directed to a singular form.

It should also be appreciated that all the modules in the apparatuses described above may be implemented in various approaches. These modules may be implemented as hardware, software, or a combination thereof. Moreover, any of these modules may be further functionally divided into sub-modules or combined together.

Processors have been described in connection with various apparatuses and methods. These processors may be implemented using electronic hardware, computer software, or any combination thereof. Whether such processors are implemented as hardware or software will depend upon the particular application and overall design constraints imposed on the system. By way of example, a processor, any portion of a processor, or any combination of processors presented in the present disclosure may be implemented with a microprocessor, microcontroller, digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic device (PLD), a state machine, gated logic, discrete hardware circuits, and other suitable processing components configured for performing the various functions described throughout the present disclosure. The functionality of a processor, any portion of a processor, or any combination of processors presented in the present disclosure may be implemented with software being executed by a microprocessor, microcontroller, DSP, or other suitable platform.

Software shall be construed broadly to mean instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, threads of execution, procedures, functions, etc. The software may reside on a computer-readable medium. A computer-readable medium may include, by way of example, memory such as a magnetic storage device (e.g., hard disk, floppy disk, magnetic strip), an optical disk, a smart card, a flash memory device, random access memory (RAM), read only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), a register, or a removable disk. Although memory is shown separate from the processors in the various aspects presented throughout the present disclosure, the memory may be internal to the processors, e.g., cache or register. The previous description is provided to enable any person skilled in the art to practice the various aspects described herein. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other aspects. Thus, the claims are not intended to be limited to the aspects shown herein. All structural and functional equivalents to the elements of the various aspects described throughout the present disclosure that are known or later come to be known to those of ordinary skilled in the art are expressly incorporated herein and intended to be encompassed by the claims.

Claims

1. A method for content recommendation based on graph-enhanced collaborative filtering, comprising: generating a candidate content item representation of a candidate content item; generating a set of historical content item representations corresponding to a set of historical content items of a target user; generating a set of general interest representations for all users based on a set of metainterests, each meta-interest representing an interest element; generating a user interest representation of the target user based on the set of historical content item representations and the set of general interest representations; and predicting a click probability of the target user clicking the candidate content item based on the candidate content item representation and the user interest representation.

2. The method of claim 1, wherein the generating a candidate content item representation comprises: generating the candidate content item representation based on a knowledge graph corresponding to the candidate content item.

3. The method of claim 2, wherein the generating the candidate content item representation comprises: identifying a set of neighbor nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighbor nodes from the knowledge graph; and generating the candidate content item representation based on at least one of: a set of neighbor node representations corresponding to the set of neighbor nodes, a set of relation representations corresponding to the set of edges, and importance of the set of neighbor nodes to the candidate content item.

4. The method of claim 1, wherein the generating a user interest representation comprises: generating the user interest representation through aggregating the set of historical content item representations with the set of general interest representations.

5. The method of claim 1, wherein the method is performed through a collaborative filtering model.

6. The method of claim 5, wherein a training of the collaborative filtering model comprises: reducing dimensionality of the set of general interest representations, to obtain a set of reduced-dimensionality general interest representations; and applying distance correlation constraints to the set of reduced-dimensionality general interest representations, to enhance a diversity of the set of reduced-dimensionality general interest representations.

26

7. The method of claim 6, wherein the reducing dimensionality of the set of general interest representations comprises: reducing the dimensionality of the set of general interest representations with Principal Component Analysis.

8. The method of claim 5, wherein a training of the collaborative filtering model comprises: training the collaborative filtering model with a training dataset including a plurality of positive content item samples and a plurality of negative content item samples, wherein the plurality of negative content item samples are sampled from a set of candidate content item samples based on a popularity of each sample in the set of candidate content item samples.

9. The method of claim 5, wherein a training of the collaborative filtering model comprises: training the collaborative filtering model with a content-based filtering model.

10. The method of claim 9, wherein the training the collaborative filtering model with a content-based filtering model comprises: generating, through the collaborative filtering model and the content-based filtering model, a cross-system contrastive prediction loss employing cross-system contrastive learning; and optimizing the collaborative filtering model at least through minimizing the cross-system contrastive prediction loss.

11. The method of claim 10, wherein the generating a cross-system contrastive prediction loss comprises: for each training data of multiple training data included in a training dataset, generating a cross-system contrastive sub prediction loss for the training data, to obtain multiple cross-system contrastive sub prediction losses for the multiple training data; and combining the multiple cross-system contrastive sub prediction losses into the cross-system contrastive prediction loss.

12. The method of claim 11, wherein the training data includes a user sample, and a positive content item sample and a negative content item sample associated with the user sample, and the generating a cross-system contrastive sub prediction loss comprises: generating, through the collaborative filtering model, a first user sample interest representation of the user sample, a first positive content item sample representation of the positive content item sample, and a first negative content item sample representation of the negative content item sample; generating, through the content-based filtering model, a second user sample interest representation of the user sample, a second positive content item sample representation of the positive content item sample, and a second negative content item sample of the negative content item sample; and generating the cross-system contrastive sub prediction loss based on the first user sample interest representation, the first positive content item sample representation, the first negative content item sample representation, the second user sample interest representation, the second positive content item sample representation and the second negative content item sample representation.

13. The method of claim 1, wherein the candidate content item or the historical content item includes at least one of movie, book, music, video, product information, and news.

14. An apparatus for content recommendation based on graph-enhanced collaborative filtering, comprising: at least one processor; and a memory storing computer-executable instructions that, when executed, cause the at least one processor to: generate a candidate content item representation of a candidate content item, generate a set of historical content item representations corresponding to a set of historical content items of a target user, generate a set of general interest representations for all users based on a set of metainterests, each meta-interest representing an interest element, generate a user interest representation of the target user based on the set of historical content item representations and the set of general interest representations, and predict a click probability of the target user clicking the candidate content item based on the candidate content item representation and the user interest representation.

15. A computer program product for content recommendation based on graph-enhanced collaborative filtering, comprising a computer program that is executed by at least one processor for: generating a candidate content item representation of a candidate content item; generating a set of historical content item representations corresponding to a set of historical content items of a target user; generating a set of general interest representations for all users based on a set of metainterests, each meta-interest representing an interest element; generating a user interest representation of the target user based on the set of historical content item representations and the set of general interest representations; and predicting a click probability of the target user clicking the candidate content item based on the candidate content item representation and the user interest representation.