CN116340645A

CN116340645A - Content recommendation based on graph-enhanced collaborative filtering

Info

Publication number: CN116340645A
Application number: CN202111587928.0A
Authority: CN
Inventors: 寿林钧; 公明; 姜大昕; 柳书广
Original assignee: Microsoft Technology Licensing LLC
Current assignee: Microsoft Technology Licensing LLC
Priority date: 2021-12-23
Filing date: 2021-12-23
Publication date: 2023-06-27
Also published as: WO2023121736A1

Abstract

The present disclosure presents methods, apparatus, and computer program products for graph-based enhanced collaborative filtering of content recommendations. Candidate content item representations of candidate content items 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 A set of historical content item representations corresponding to a set of historical content items of a target user may be generated.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 overall interest representations for the population of users may be generated based on a set of meta-interests, each meta-interest characterizing an element of interest. A user interest representation of the target user may be generated based on the set of historical content item representations and the set of overall interest representations. A click probability of the target user clicking on 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 play an increasingly important role in many online services. There are different recommendation systems, such as a movie recommendation system, a book recommendation system, a music recommendation system, a product recommendation system, and the like, based on the recommended content. These recommendation systems typically capture the interests of the user and predict and recommend content of interest to the user based on the user interests.

Disclosure of Invention

This summary is provided to introduce a selection of concepts that are further described below in the detailed description. This summary 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 methods, apparatuses and computer program products for graph-based enhanced collaborative filtering of content recommendations. Candidate content item representations of candidate content items 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 overall interest representations for the population of users may be generated based on a set of meta-interests, each meta-interest characterizing an element of interest. A user interest representation of the target user may be generated based on the set of historical content item representations and the set of overall interest representations. A click probability of the target user clicking on the candidate content item may be predicted based on the candidate content item representation and the user interest representation.

It is noted that one or more of the aspects above include the features specifically pointed out in the following detailed description and the claims. The following description and the annexed drawings set forth in detail certain illustrative features of the one or more aspects. These features are indicative of but a few of the various ways in which the principles of various aspects may be employed and the present disclosure is intended to include all such aspects and their equivalents.

Drawings

The disclosed aspects will be described below in conjunction with the drawings, which are provided to illustrate and not limit the disclosed aspects.

FIG. 1 illustrates an exemplary process for graph-based enhanced collaborative filtering of content recommendations in accordance with an embodiment of the present disclosure.

FIG. 2 illustrates an exemplary process for training collaborative filtering models according to an embodiment of the present disclosure.

FIG. 3 illustrates an exemplary process of employing a content-based filtering model to predict click probabilities in accordance with 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 disclosure.

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

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

FIG. 7 is a flowchart of an exemplary method for graph-based enhanced collaborative filtered content recommendation, in accordance with an embodiment of the present disclosure.

FIG. 8 illustrates an exemplary apparatus for graph-based enhanced collaborative filtering of content recommendations in accordance with an embodiment of the present disclosure.

FIG. 9 illustrates an exemplary apparatus for graph-based enhanced collaborative filtering of content recommendations in accordance with an embodiment of the present disclosure.

Detailed Description

The present disclosure will now be discussed with reference to several exemplary embodiments. It should be understood that the discussion of these embodiments is merely intended to enable one skilled in the art to better understand and thereby practice the examples of the present disclosure and is not intended to limit the scope of the present disclosure in any way.

Collaborative filtering (Collaborative Filtering, CF) is a widely used recommended technique. Collaborative filtering techniques aim to identify similar users that are similar to a target user, determine interests of the target user through interests of the similar users, and make content recommendations to the target user based on the determined interests of the target user. In this context, the target user may refer to the user for which content recommendation is performed. The collaborative filtering techniques described above may be implemented by a machine learning model. The machine learning model employing the collaborative filtering-based content recommendation method may be referred to herein as a collaborative filtering model. The click probability of the target user clicking on each candidate content item in the set of candidate content items may be predicted by a collaborative filtering model to obtain a set of click probabilities. In this context, a content item may refer to an individual item having specific content. For example, a movie, a book, a piece of music, etc. may be referred to as a content item. The content items to be recommended to the target user may be determined by ordering the set of click probabilities.

The collaborative filtering model may generate a candidate content item representation of the candidate content item and a user interest representation of the target user and predict a click probability of the target user clicking on the candidate content item based on the candidate content item representation and the user interest representation. In general, historical content items previously browsed, clicked or viewed by the target user may indicate the user interests of the target user, and thus a user interest representation of the target user may be generated based on representations of the historical content items. In generating representations of content items such as candidate content items and historical content items, a Knowledge Graph (knowledgegraph) corresponding to the content item may be considered. The knowledge graph may contain a number of nodes representing a number of entities and a number of edges representing rich relationships between the entities. In the knowledge graph, a neighbor node adjacent to a particular node may be considered as structural information of the particular node, such as attribute information, and an edge from the neighbor node to the particular node may be considered as a relationship of the neighbor node with respect to the particular node. Therefore, when collaborative filtering technology is adopted to conduct content recommendation, the candidate content item representation and the user interest representation can be enriched by utilizing the knowledge graph, so that the click probability of a target user clicking on the candidate content item can be predicted more accurately. A method of performing content recommendation using knowledge-graph-based collaborative filtering may be referred to as a graph-enhanced collaborative filtering (Graph Enhanced Collaborative Filtering) -based content recommendation method.

Embodiments of the present disclosure propose improved graph-based enhanced collaborative filtering of content recommendations. Candidate content item representations of the candidate content items and user interest representations of the target user may be generated, and a click probability of the target user clicking on the candidate content items may be predicted based on the generated candidate content item representations and user interest representations. In generating the user interest representation, a set of historical content item representations corresponding to a set of historical content items of the target user and a set of overall interest representations for the overall user may be generated, and the user interest representation may be generated based on the set of historical content item representations and the set of overall interest representations. The general interest representation may be corresponding to the general interest. In this context, the overall interest may refer to the interest of the entire user, not of a particular user. Embodiments of the present disclosure propose that a set of overall interest representations may be generated based on a set of element interests. In this context, meta-interests may refer to one of a set of interest elements that constitute an overall interest for a population of users. Each metainterest may have a corresponding trainable embedded vector. Generating a general interest representation using meta-interests may more broadly capture different points of interest in the general interests for the population of users. The above improved graph-enhanced collaborative filtering-based content recommendation method may be performed by a collaborative filtering model according to an embodiment of the present disclosure.

In one aspect, embodiments of the present disclosure propose to generate candidate content item representations of candidate content items based on knowledge maps corresponding to the candidate content items. For example, a set of neighbor nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighbor nodes may be identified from the knowledge graph, and the candidate content item representation may be generated based on the set of neighbor node representations corresponding to the set of neighbor nodes, the set of relationship representations corresponding to the set of edges, the importance of the set of neighbor nodes to the candidate content item, and so on. The importance of the candidate content item by the neighbor node may be used to control the information transferred from the neighbor node to the candidate content item. Neighbor nodes that are more important to the candidate content item may communicate more information to the candidate content item and thus may affect the candidate content item representation to a greater extent. Similarly, for each historical content item in a 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, embodiments of the present disclosure propose to employ soft distance correlations (Soft Distance Correlation) to enhance the diversity of a set of overall interest representations used to generate a user interest representation when training collaborative filtering models. For example, a set of overall interest representations may be reduced in dimension to obtain a set of reduced-dimension overall interest representations, and distance-related constraints may be applied to the set of reduced-dimension overall interest representations to enhance diversity of the set of reduced-dimension overall interest representations. Applying a distance-dependent constraint to a set of general interest representations may cause the individual general interest representations to be separated from each other in vector space, thus ensuring that each general interest representation is as different as possible, so that different general interest representations may characterize different interests from different aspects. Further, first reducing the dimensions of a set of overall interest representations and applying distance-related constraints to only a set of reduced-dimension overall interest representations may ensure diversity of the overall interest representations at low dimensions while maintaining flexibility at high dimensions. In this way, different points of interest of the user may be more fully learned, such that a more accurate representation of the user's interests may be generated.

In yet another aspect, embodiments of the present disclosure propose to determine a negative content item sample by inverse ratio negative sampling (Reciprocal Ratio Negative Sampling) when constructing a training dataset for training a collaborative filtering model. In this context, a negative content item sample may refer to a content item sample that is not of interest to the user. Negative content item samples may be sampled from a set of candidate content item samples based on popularity of each content item sample in the set. For example, content items with more user interactions may be considered popular content items. Popular content items may have a smaller probability of being sampled as negative content item samples. Unlike existing approaches that determine negative content item samples by random sampling, determining negative content item samples by inverse ratio negative sampling may avoid determining content items that a user has not previously encountered but that may be of interest as negative content item samples for the user, and may thereby improve the quality of the negative content item samples.

In yet another aspect, embodiments of the present disclosure propose to train collaborative Filtering models using Content-based Filtering (CBF) models. The content-based filtering model may utilize unstructured information of the candidate content item and/or the historical content item, such as a textual description of the content embodying the candidate content item and/or the historical content item, in generating a representation of the candidate content item and/or the historical content item. The historical content item representations may be further used to generate a user interest representation of the target user. Training the collaborative filtering model with the content-based filtering model may cause knowledge learned by the content-based filtering model, such as knowledge obtained from unstructured information of the content item, to be migrated to the collaborative filtering model. Collaborative filtering models may be trained using content-based filtering models through Cross-system contrast learning (Cross-System Contrastive Learning). This approach provides a lightweight way to blend collaborative filtering models and content-based filtering models. Collaborative filtering models trained using content-based filtering models may take into account textual descriptions of candidate content items and/or historical content items in generating candidate content item representations and/or user interest representations, such that richer candidate content item representations and/or user interest representations may be generated. In addition, for some new content items, such as newly online movies, newly published books, etc., recommendations can be made based on the text description of the content item, even if there is less user interaction for it or it is not involved in previous training, which can effectively solve the cold start problem.

FIG. 1 illustrates an exemplary process 100 for graph-based enhanced collaborative filtering of content recommendations in accordance with embodiments of the present disclosure. In process 100, target user u-click candidate content item 102i (i e S) may be predicted by collaborative filtering model 110 _I ) Click probability 112 of (c).

The candidate content item 102i may be a content item from a group of candidate content items that may be recommended to the target user u. Candidate content items 102i may include, for example, movies, books, music, video, product information, news, and the like.

Candidate content item representations 122 for the candidate content item 102i may be generated by the content item encoder 120. The content item encoder 120 may be, for example, a gated path graph convolution network (Gated Path Graph Convolution Network) comprising L convolution layers. Preferably, the content item encoder 120 may generate candidate content item representations 122 based on the knowledge-graph 106. Knowledge graph 106 may be labeled as

Wherein->

Is a set of nodes in the knowledge-graph, and ε is a set of edges between the set of nodes. The knowledge-graph 106 may be a knowledge-graph corresponding to the candidate content item 102 i. Taking the example where the candidate content item 102i is a movie, the knowledge-graph 106 may be a knowledge-graph that includes a set of nodes whose entity types are movies and a set of edges between the nodes. Candidate content item 102i may correspond to knowledge graph 106 +. >

Node in (a)

Node v may be identified from knowledge-graph 106 _i Adjacent set of neighbor nodes

And a set of edges corresponding to the set of neighbor nodes +.>

Convolution layer l+1 (0.ltoreq.l) in a gated path graph convolution network<L) may be based on node v _i Is->

Representation at convolution layer l, corresponding to node v _i And the group of neighbor nodes->

A set of relation representations of a set of edges between, the set of neighbor nodes +.>

Opposite node v _i Generating node v by importance of (a) or the like _i The representation at l+1 is as shown in the following equation:

wherein, the liquid crystal display device comprises a liquid crystal display device,

is slave node v _i To neighbor node v _j Relation r of (2) _ij Is expressed by the relation of gamma _ij Is to slave neighbor node v _j To node v _i A gating function controlling the information transferred, which can be used to control the slave neighbor node v _j Is transferred to node v _i Weighting the information of (a). />

May be node v _i Or an initial representation of the candidate content item i, which may be a randomly initialized representation or a representation obtained by a known knowledge-graph embedding vector method. Similarly, the relationship represents +.>

The representation may be a randomly initialized representation or a representation obtained by a known knowledge-graph embedded vector method.

Gating function gamma _ij Can reflect neighbor node v _j Opposite node v _i Is of importance. The gating function γ can be calculated by the following formula _ij ：

Where σ (·) is a sigmoid function for limiting the gating value between 0 and 1. Opposite node v _i The more important neighbor nodes can be towards node v _i The more information is transferred, the greater the influence on node v can be _i I.e. the candidate content item i. It should be appreciated that although in equation (1), the importance of candidate content items based on the inclusion of nodes such as neighbor node representations, relationship representations, and neighbor nodes is illustratedThree factors such as sex, but in some embodiments, it is also possible to generate candidate content item representations based on only any one or any two of these three factors.

To overcome the problem of excessive smoothing of graph convolution, node v may be aggregated _i Representation at all intermediate layers to obtain node v _i Final representation of (i) candidate content item representation 122

As shown in the following formula:

to generate the user interest representation 142 for the target user u, a set of historical content items 104 for the target user u may be obtained. The set of historical content items 104 may include a plurality of historical content items that the target user u has previously browsed, clicked or viewed, such as historical content items 104-1 through 104-C, where C is the number of historical content items. Historical content items may include, for example, movies, books, music, video, product information, news, and the like. A set of historical content items 104 for target user u may indicate the user interests of target user u. Taking the example of a movie for a historical content item, movies previously watched by the target user u may indicate which movies the user is interested in.

A set of historical content item representations 132 corresponding to a set of historical content items 104 may be generated by a set of content item encoders 130. The set of content item encoders 130 may include, for example, content item encoder 130-1 through content item encoder 130-C. The set of historical content item representations 132 may include, for example, historical content item representations 132-1 through historical content item 132-C. Each content item encoder of the set of content item encoders 130 may have a similar structure to the content item encoder 120, for example, may be a gated path graph convolution network comprising L convolution layers. The historical content item representations 132-1 through 132-C may be generated by a process similar to that of 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, such as the knowledge-graph 106, corresponding to the historical content item.

Subsequently, a user interest representation 142 of the target user u may be generated by the user encoder 140. For example, a set of overall interests 108 for the population of users may be based

To generate a set of overall interest representations for the population of users. The group general interest- >

Interactions of the totality of users with the content item may be involved. The user interest representation 142 for the target user u may be generated based on a set of historical content item representations 132 and a set of overall interest representations.

According to an embodiment of the present disclosure, the set of general interests

Is +.>

Can be made of a set of interest->

The composition is formed. Every meta-interest->

An element of interest may be characterized. For each general interest p, the interest can be made by this component +.>

To generate a global interest representation e of the global interest p _p As shown in the following formula:

wherein e _m ∈R ^h Is a trainable embedded vector of meta-interest m, and

is a slave trainable weight for the overall interest p as shown in the following formula>

Linear weights derived from:

can be aimed at each general interest

To perform the above operations to generate a set of general interest representations { e } _p }. Next { e } can be represented by utilizing a set of general interests _p Aggregation of a set of historical content item representations 132 at convolution layer l to generate a representation of user interest at convolution layer l, as shown in the following equation:

is the historical interactions of user u, +.>

Is a representation of the history content item i at the convolution layer i, and alpha _p Is a weight for overall interest obtained through the attention mechanism as shown in the following formula:

Similar to generating candidate content item representations 122, to overcome the problem of excessive smoothing of graph convolutions, a final representation of user interests, namely user interest representation 142, may be obtained by aggregating representations of user interests at all intermediate layers

As shown in the following formula:

at the time candidate content item representation 122 is generated

And user interest representation 142->

The click probability 112 of the target user u clicking on the candidate content item 102i may then be predicted by the prediction layer 150. Click probability may be marked +.>

In one embodiment, 142 may be represented by user interest>

And candidate content item representation 122->

Applying dot product operation to predict click probability +.>

As shown in the following formula:

because collaborative filtering model 110 may utilize Knowledge-graphs to generate candidate content item representations and/or historical content item representations, and may generate user Interest representations based on a set of Meta-interests, collaborative filtering model 110 may also be referred to as a Knowledge-Graph Enhanced Meta-Interest Network (knowledged-Graph-Enhanced Meta-Interest Network).

It should be appreciated that the process for graph-based enhanced collaborative filtering of content recommendations described above in connection with FIG. 1 is merely exemplary. The steps in the process for graph-based enhanced collaborative filtered content recommendation may be replaced or modified in any manner and may include more or fewer steps depending on the actual application requirements. For example, although in process 100, knowledge maps corresponding to candidate content items and/or historical content items are considered in generating candidate content item representations and/or historical content item representations, in some embodiments, it is possible to disregard knowledge maps. In this case, the respective candidate content item representations and/or historical content item representations may be generated based solely on the candidate content items and/or the historical content items themselves. Further, the particular order or hierarchy of steps in process 100 is merely exemplary, and the process for graph-based enhanced collaborative filtering of content recommendations may be performed in an order different from that described.

Collaborative filtering models, such as collaborative filtering model 110 in fig. 1, may be trained in a variety of ways. FIG. 2 illustrates an exemplary process 200 for training collaborative filtering models in accordance with embodiments of the present disclosure. The collaborative filtering model trained through process 200, when actually deployed, may predict the click probability of a target user clicking on a candidate content item.

As described in connection with fig. 1, a set of overall interest representations for the totality of users is considered in generating the user interest representation for the target user. Embodiments of the present disclosure propose to employ soft distance correlations to enhance the diversity of the set of overall interest representations.

For example, at 202, a set of overall interest representations may be reduced in dimension to obtain a set of reduced-dimension overall interest representations. In one embodiment, principal component analysis (Principal Component Analysis, PCA) may be utilized to reduce the dimensions of a set of overall interest representations, as shown in the following equation:

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

At 204, a soft distance-dependent predictive loss may be generated by applying a distance-dependent constraint (distance correlation constraint) to the set of reduced-dimension global interest representations

As shown in the following formula:

wherein DCov (·) is used to calculate the distance covariance and DVar (·) is used to calculate the distance variance. It should be appreciated that, when e=1 in equation (10),

the result is the original distance-dependent predictive loss.

Applying a distance-dependent constraint to a set of general interest representations may cause the individual general interest representations to be separated from each other in vector space, thus ensuring that each general interest representation is as different as possible, so that different general interest representations may characterize different interests from different aspects. Further, first reducing the dimensions of a set of overall interest representations and applying distance-related constraints to only a set of reduced-dimension overall interest representations may ensure diversity of the overall interest representations at low dimensions while maintaining flexibility at high dimensions. In this way, different points of interest of the user may be more fully learned, such that a more accurate representation of the user's interests may be generated.

Additionally or alternatively, the collaborative filtering model may be trained with a training data set comprising 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 that user u in a set of candidate content item samples previously browsed, clicked or viewed may be considered as a positive content item sample i for user u ⁺ 。

At 212, a plurality of negative content item samples { i } may be obtained ^- }. Embodiments of the present disclosure propose to determine a negative content item sample i by inverse 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 popularity of each content item sample in the set ^- }. For example, content items with more user interactions may be considered popular content items. Popular content items may have a smaller probability of being sampled as a negative content item sample i ^- . Sample i e S for content item _I Which is sampled as a negative content item sample i ^- The probability of (2) may be expressed as:

where c (i) represents a count of the interactions of the totality of users with the content item sample i. Unlike existing approaches that determine negative content item samples by random sampling, determining negative content item samples by inverse ratio negative sampling may avoid determining content items that a user has not previously encountered but that may be of interest as negative content item samples for the user, and may thereby improve the quality of the negative content item samples.

At 214, a plurality of positive content item samples { i } may be based ⁺ Sum of a plurality of negative content item samples { i } ^- Building training data set

At the point of 216 the process continues with the step of,paired bayesian personalized ordering (Bayesian Personalized Ranking, BPR) predictive losses can be generated based on the constructed dataset

The BPR predictive loss aims to assign a higher score to content items that the user has browsed than content items that the user is not interested in using the concept of contrast learning. The BPR predictive loss +_ can be generated, for example, by the following formula>

Alternatively or additionally, at 218, an L2 regularization (regularization) predictive loss may be generated based on the constructed data set

Wherein->

And->

Is the L2 norm of the user interest representation/content item representation.

Subsequently, at 220, a composite predicted loss of the collaborative filtering model may be generated based on at least one of the distance-related predicted loss, the BPR predicted loss, and the L2 regularized predicted loss

As shown in the following formula:

wherein lambda is ₁ And lambda (lambda) ₂ Is used for controlling predictionSuper parameters of the loss weights.

At 222, the loss may be predicted by integrating

Minimization is used to train collaborative filtering models.

It should be appreciated that the process for training the collaborative filtering model described above in connection with FIG. 2 is merely exemplary. The steps in the process for training the collaborative filtering model may be replaced or modified in any manner depending on the actual application requirements, and the process may include more or fewer steps. For example, although in process 200, three distance-dependent predictive losses, BPR predictive losses, and L2 regularized predictive losses are considered in generating the integrated predictive losses of the collaborative filtering model, in some embodiments, the integrated predictive losses of the collaborative filtering model may be generated based only on any one or any two of the distance-dependent predictive losses, BPR predictive losses, and L2 regularized predictive losses. Further, the particular order or hierarchy of steps in process 200 is merely exemplary, and the process for training collaborative filtering models may be performed in an order different from that described.

Generating a candidate content item representation of the candidate content item and a user interest representation of the target user by a collaborative filtering model, and then predicting a click probability of the target user clicking on the candidate content item based on the generated candidate content item representation and the user interest representation, is described above in connection with fig. 1. The collaborative filtering model considers structured information of content items, such as various types of attribute information from knowledge maps, in generating candidate content item representations and/or user interest representations. Embodiments of the present disclosure propose to train collaborative filtering models using content-based filtering models. Similar to the collaborative filtering model, the content-based filtering model may generate candidate content item representations of candidate content items and user interest representations of the target user, and predict a click probability of the target user clicking on the candidate content item based on the candidate content item representations and the user interest representations. Furthermore, the user interest representation may also be generated based on representations of historical content items of the user. However, the content-based filtering model may utilize unstructured information of the candidate content item and/or the historical content item, such as a textual description of the content embodying the candidate content item and/or the historical content item, in generating a representation of the candidate content item and/or the historical content item. Collaborative filtering models may be trained by cross-system contrast learning using content-based filtering models, which provides a lightweight way to blend collaborative filtering models and content-based filtering models. The text descriptions of the candidate content items and/or the historical content items may be taken into account in generating candidate content item representations and/or user interest representations using collaborative filtering models trained on content-based filtering models, such that richer candidate content item representations and/or user interest representations may be generated. In addition, for some new content items, such as newly online movies, newly published books, etc., recommendations can be made based on the text description of the content item, even if there is less user interaction for it or it is not involved in previous training, which can effectively solve the cold start problem.

FIG. 3 illustrates an exemplary process 300 for predicting click probabilities using a content-based filtering model in accordance with an embodiment of the disclosure. In process 300, the click probability 312 of the target user u clicking on the candidate content item 302i may be predicted by a content-based filtering model 310.

Candidate content item 302i may be a content item from a set of candidate content items that may be recommended to target user u. Candidate content items 302i may include, for example, movies, books, music, video, product information, news, and the like. A textual description 306x of the candidate content item 302i may be obtained _i . The textual description 306x may be obtained in a variety of known ways _i . In one embodiment, the textual description 306x may be extracted from a knowledge-graph corresponding to the candidate content item 302i _i . In another embodiment, articles in the Internet that are related to the candidate content item 302i may be searched for by a trained machine learning model, and important paragraphs are identified from the searched articles as text descriptions 306x of the candidate content item 302i _i 。

Text-based description 306x may be passed through content item encoder 320 _i To generate candidate content item representations 322e of candidate content item 302i _i . The content item encoder 320 may be, for example, a Pre-trained converter-based model, such as a bi-directional encoder representation (Bidirectional Encoder Representations from Transformers, BERT) model from a converter, a Generative Pre-trained Transformer-2, gpt-2 model, or the like. Taking the example where the content item encoder 320 is a BERT model, the candidate content item representations 322e _i The following formula may be used:

a set of historical content items 304 for the target user u may be obtained. The set of historical content items 304 may include a plurality of historical content items that the target user u has previously browsed, clicked, or viewed, such as historical content items 304-1 through historical content item 304-B, where B is the number of historical content items. Historical content items may include, for example, movies, books, music, video, product information, news, and the like. A set of historical content items 304 for target user u may indicate the user interests of target user u. A set of text descriptions 308 corresponding to a set of historical content items 304 may be obtained. For example, a textual description 308-1 of the historical content item 304-1 may be obtained, a textual description 308-B of the historical content item 304-B may be obtained, and so on.

A set of historical content item representations 332 corresponding to a set of historical content items 304 may be generated by a set of content item encoders 330. The set of content item encoders 330 may include, for example, content item encoder 330-1 through content item encoder 330-B. The set of historical content item representations 332 may include, for example, historical content item representations 332-1 through 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, and may be, for example, a pre-trained converter-based model such as a BERT model, a GPT-2 model, or the like. The historical content item representations 332-1 through 332-B may be generated by a process similar to that of the candidate content item representations 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 textual description of the historical content item.

Subsequently, a user interest representation 342e of the target user u may be generated by the user encoder 340 _u . The user encoder 340 may generate the user interest representation 342e based on the set of historical content item representations 332 for the target user u _u . In one implementation, the user interest representation 342e may be generated by weighted summing the B historical content item representations included in the set of historical content item representations 332 _u As shown in the following formula:

wherein alpha is _b Is assigned to the history content item i _u,b Is obtained by passing features through two linear layers, as shown in the following formula:

and->

The weights and offsets of the two fully connected layers, respectively.

At the time candidate content item representation 322e is generated _i And user interest representation 342e _u The click probability 312 of the target user u clicking on the candidate content item 302 may then be predicted by the prediction layer 350. Clicking on the summaryThe rate may be marked as

In one embodiment, the representation 342e may be represented by interest to the user _u And candidate content item representation 322e _i Applying dot product operation to predict click probability +.>

As shown in the following formula:

since the content-based filtering model 310 employs a multi-headed Self-Attention mechanism, it may also be referred to as a Neural Recommendation (NRMS) model with multi-headed Self-Attention. In particular, where the content-based filtering model 310 employs a pre-trained BERT model as the 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 of employing a content-based filtering model to predict click probabilities described above in connection with FIG. 3 is merely exemplary. The steps in the process of predicting click probabilities using the content-based filtering model may be replaced or modified in any manner and may include more or fewer steps depending on the actual application requirements. Further, the particular order or hierarchy of steps in process 300 is merely exemplary, and the process of employing a content-based filtering model to predict click probabilities may be performed in an order different from the order described.

In accordance with embodiments of the present disclosure, collaborative filtering models may be trained using content-based filtering models. For example, collaborative filtering model 110 in FIG. 1 may be trained using content-based filtering model 310 in FIG. 3. Preferably, the content-based filter model may be pre-trained prior to training the collaborative filter model with the content-based filter model. A negative sampling method (negative sampling method) may be employed to train the content-based filtering model. Fig. 4 illustrates an exemplary process 400 for training a content-based filtering model in accordance with an embodiment of the disclosure. The content-based filtering model trained through process 400, when actually deployed, may predict the click probability of a target user clicking on a candidate content item.

First, a training data set for training a content-based filtering model may be constructed. In one embodiment, a list-wise policy (list-wise structy) may be employed to construct the training data set. For example, at 410, a plurality of positive content item samples may be obtained. For example, content items in a set of candidate content item samples that have been previously viewed, clicked on, or viewed by a user may be considered 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 set of content item samples that are presented in the same session but not clicked on by the user may be considered as a negative set of content item samples corresponding to the positive content item sample.

At 430, a training data set 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, positive content item sample click probabilities corresponding to each positive content item sample may be predicted. The positive content item sample click probability corresponding to the ith positive content item sample may be marked as

At 450, for each negative content item sample in the 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 setAnd (5) combining. The negative content item sample click probability set corresponding to the negative content item sample set of the ith positive content item sample may be labeled as

Where K is the number of negative content item samples comprised by the negative content item sample click probability set. In this way, the click probability prediction problem can be expressed as a pseudo K+1-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. The posterior click probability corresponding to the ith positive content item sample may be labeled p _i . In one implementation, the positive content item sample click probability may be determined by using a softmax function

And negative content item sample click probability set +.>

Normalization is performed to calculate the posterior click probability corresponding to the positive content item sample, as shown in the following equation:

The operations at steps 440 through 460 described above may be performed for each of a plurality of positive content item samples in the training data set such that at 470, a plurality of posterior click probabilities corresponding to the plurality of positive content item samples may be obtained.

At 480, a predictive loss may be generated based on the plurality of posterior click probabilities. In one embodiment, the predictive loss may be generated by calculating a negative log-likelihood (negative log-likelihood) of a plurality of posterior click probabilities, as shown in the following equation:

wherein S is _i Is a positive content item sample set consisting of a plurality of positive content item samples.

At 490, the content-based filtering model may be optimized by minimizing predictive loss.

It should be appreciated that the process for training the content-based filtering model described above in connection with FIG. 4 is merely exemplary. The steps in the process for training the content-based filtering model may be replaced or modified in any manner and may include more or fewer steps depending on the actual application requirements. For example, while in process 400 a tabular strategy is employed to construct the training data set, other ways of constructing the training data set are possible, such as the inverse ratio negative sampling method described in connection with FIG. 2, in determining the negative content item samples. Further, the particular order or hierarchy of steps in process 400 is merely exemplary, and the process for training the content-based filtering model may be performed in an order different from that described.

Embodiments of the present disclosure propose to train collaborative filtering models with content-based filtering models through cross-system contrast learning. FIG. 5 illustrates another exemplary process 500 for training collaborative filtering models according to an embodiment of the present disclosure. In process 500, cross-system contrast prediction losses may be generated by employing cross-system contrast learning through collaborative filtering models and content-based filtering models in addition to generating a composite prediction loss of the collaborative filtering model based on at least one of soft distance-related prediction losses, BPR prediction losses, and L2 regularized prediction losses, as compared to process 200 described in connection with fig. 2.

Steps 502 to 520 may correspond to steps 202 to 220, respectively, in fig. 2.

At 530, cross-system contrast prediction loss may be generated using cross-system contrast learning through collaborative filtering models and content-based filtering models

The generation of cross-system contrast prediction loss will be described later in connection with FIG. 6>

Is described herein).

At 540, loss may be predicted based on the synthesis

And cross-system contrast prediction loss->

Cross-system enhanced comprehensive predictive loss ++generating collaborative filtering model>

As shown in the following formula:

Wherein lambda is _cs For controlling cross-system contrast prediction loss

Is a weight of (2).

At 550, the loss may be predicted by having the integrated across-system enhanced

Minimization optimizes collaborative filtering models.

It should be appreciated that the process for training a collaborative filtering model using a content-based filtering model described above in connection with FIG. 5 is merely exemplary. The steps in the process for training collaborative filtering models with content-based filtering models may be replaced or modified in any manner depending on the actual application requirements, 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 integrated predicted loss and the cross-system contrast predicted loss, in some embodiments, the collaborative filtering model may be trained based only on the cross-system contrast predicted loss. In this case, the collaborative filtering model may be optimized by minimizing cross-system contrast prediction loss. Further, the particular order or hierarchy of steps in process 500 is merely exemplary, and the process for training collaborative filtering models using content-based filtering models may be performed in an order different from the order described.

FIG. 6 illustrates an exemplary process 600 for generating cross-system contrast prediction losses in accordance with embodiments of the present disclosure. Process 600 may correspond to step 530 in fig. 5.

The pre-constructed training data set 610 may be utilized to generate cross-system contrast prediction loss 660

Training data set 610 may be, for example, a training data set constructed by steps 510 to 514 in fig. 5, i.e., steps 210 to 214 in fig. 2 ∈>

For training data set 610->

Each training data 620 d= (u, i) ⁺ ,i ^- ) Cross-system versus contrast prediction loss 650 ++may be generated for this training data 620d>

Training data 620d may include, for example, user samples 622u, positive content item samples 624i ⁺ And negative content item sample 626i ^- . Positive content item sample 624i ⁺ And negative content item sample 626i ^- May be associated with user sample 622 u. For example, positiveContent item sample 624i ⁺ May be a content item viewed by user sample 622u, while negative content item sample 626i ^- May be a content item that is not of interest to the user sample 622 u.

Collaborative filtering model 630 may correspond, for example, to collaborative filtering model 110 in fig. 1. A first user sample interest representation 632 of the user sample 622u may be generated by collaborative filtering model 630

Positive content item sample 624i ⁺ Is the first positive content item sample representation 634 +.>

And negative content item sample 626i ^- Is 636 +.>

The content-based filter model 640 may correspond, for example, to the content-based filter model 310 in fig. 3. A second user sample interest representation 642 of the user sample 622u may be generated by the content-based filtering model 640

Positive content item sample 624i ⁺ Is a second positive content item sample representation 644 +.>

And negative content item sample 626i ^- Is a second negative content item sample representation 646 +.>

The content-based filter model 640 may be based on positive content item samples 624i ⁺ Generating a second positive content item sample representation 644 +.>

Similarly, the content-based filter model 640 may be based on negative content item samples626i ^- To generate a second negative content item sample representation 646 +.>

Subsequently, a representation 632 may be represented based on the first user sample interest

The first positive content item sample representation 634 +.>

The first negative content item sample representation 636 +.>

Second user sample interest representation 642->

Second positive content item sample representation 644>

And a second negative content item sample representation 646 ∈ ->

To generate cross-system versus sub-predictive loss 650 +/for training data 620d>

For example, cross-system contrast learning may be employed to generate cross-system contrast predicted loss 650 +. >

As shown in the following formula:

may be directed to a training data set

Each of the plurality of training data in (1) performs the above operation to obtain a plurality of cross-system contrast sub-prediction losses +.>

These multiple cross-system vs. sub-predictive losses +.>

May be combined into a cross-system contrast predicted loss 660 +_ for collaborative filtering model 630>

As shown in the following formula:

cross-system contrast prediction loss

The collaborative filtering model may be directed to fuse content-sensitive information, such as information related to textual descriptions, from the content-based filtering model. For example, cross-system contrast prediction loss +.>

The collaborative filtering model may be directed to generate a content item representation that interacts not only with the user interest representation output by the collaborative filtering model itself, but also with the user interest representation output by the content-based filtering model. Similarly, cross-system contrast prediction loss +.>

The collaborative filtering model may be directed to generate a user-interest representation that interacts not only with the content item representations output by the collaborative filtering model itself, but also with the content item representations output by the content-based filtering model. This way collaborative filtering models can be guided in generating contentThe item representations and/or the user interest representations are time-wise learned to learn representations of textual descriptions of the content items that are output by the content-based filtering model. Thus, knowledge learned by the content-based filtering model may be migrated to the collaborative filtering model.

It should be appreciated that the process for generating cross-system contrast prediction losses described above in connection with fig. 6 is merely exemplary. The steps in the process for generating cross-system contrast prediction loss may be replaced or modified in any manner and may include more or fewer steps depending on the actual application requirements. Further, the particular order or hierarchy of steps in process 600 is exemplary only, and the process for generating cross-system contrast prediction loss may be performed in an order different from that described.

Content recommendation based on graph-enhanced collaborative filtering according to embodiments of the present disclosure may be applicable to various types of recommendations, such as movie recommendation, book recommendation, music recommendation, video recommendation, product recommendation, news recommendation, and the like. In particular, when a collaborative filtering model according to an embodiment of the present disclosure, such as collaborative filtering model 110 shown in fig. 1, is used for online deployment, it is particularly suitable for relatively static domain recommendations such as movie recommendations, book recommendations, music recommendations, video recommendations, product recommendations, etc. that are slow to update content or have more user interactions, since the collaborative filtering model fully accounts for user interactions. When a content-based filtering model, such as content-based filtering model 310 shown in fig. 3, according to embodiments of the present disclosure is used for online deployment, it is particularly suitable for relatively dynamic domain recommendations, such as news recommendations, that update faster or have less user interaction for content, as the content-based filtering model may be recommended based on text information of content items.

FIG. 7 is a flowchart of an exemplary method 700 for graph-based enhanced collaborative filtered content recommendation, in accordance with an embodiment of the present disclosure.

At 710, candidate content item representations of candidate content items may be generated.

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

At 730, a set of overall interest representations for the population of users may be generated based on a set of meta-interests, each meta-interest characterizing an element of interest.

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 overall interest representations.

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

In one implementation, the generating the candidate content item representations may include: the candidate content item representations are generated based on a knowledge-graph corresponding to the candidate content item.

The generating the candidate content item representation may include: identifying, from the knowledge-graph, a set of neighbor nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighbor nodes; 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 relationship representations corresponding to the set of edges, and importance of the set of neighbor nodes to the candidate content item.

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

In one embodiment, the generating the user interest representation may include: the user interest representation is generated by aggregating the set of historical content item representations with the set of overall interest representations.

In one embodiment, the method 700 may be performed by collaborative filtering models.

The training of the collaborative filtering model may include: reducing the dimensions of the set of overall interest representations to obtain a set of reduced-dimension overall interest representations; and applying a distance-related constraint to the set of reduced-dimension global interest representations to enhance diversity of the set of reduced-dimension global interest representations.

The reducing the dimension of the set of overall interest representations may include: principal component analysis is utilized to reduce the dimensions of the set of overall interest representations.

The training of the collaborative filtering model may include: the collaborative filtering model is trained using a training data set comprising 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 popularity of each content item sample in the set.

The training of the collaborative filtering model may include: the collaborative filtering model is trained using a content-based filtering model.

The training the collaborative filtering model with the content-based filtering model may include: generating cross-system contrast prediction loss by adopting cross-system contrast learning through the collaborative filtering model and the content-based filtering model; and optimizing the collaborative filtering model at least by minimizing the cross-system contrast prediction loss.

The generating cross-system contrast prediction loss may include: for each training data of a plurality of training data included in a training data set, generating cross-system versus sub-prediction losses for the training data to obtain a plurality of cross-system versus sub-prediction losses for the plurality of training data; and combining the plurality of cross-system contrast sub-predictive losses into the cross-system contrast predictive losses.

The training data may include a user sample, a positive content item sample and a negative content item sample associated with the user sample. The generating cross-system versus sub-prediction loss may include: generating, by 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, by 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 representation of the negative content item sample; and generating the cross-system contrast sub-prediction penalty 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 the second positive content item sample representation may comprise: the second positive content item sample representation is generated based on the text description of the positive content item sample by the content-based filtering model. The generating the second negative content item sample representation may comprise: the second negative content item sample representation is generated based on the text description of the negative content item sample by the content-based filtering model.

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

It should be appreciated that method 700 may also include any steps/processes for graph-based enhanced collaborative filtering of content recommendations according to embodiments of the present disclosure described above.

FIG. 8 illustrates an example apparatus 800 for graph-based enhanced collaborative filtering of content recommendations in accordance with an embodiment of the present disclosure.

The apparatus 800 may include: a candidate content item representation generation module 810 for generating candidate content item representations of candidate content items; a historical content item representation generation 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 generation module 830 for generating a set of general interest representations for the population of users based on a set of meta-interests, each meta-interest characterizing an interest element; a user interest representation generation 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 overall interest representations; and a click probability prediction module 850 for predicting a click probability of the target user clicking on the candidate content item based on the candidate content item representation and the user interest representation. In addition, apparatus 800 may also include any other modules configured for graph-based enhanced collaborative filtering of content recommendations in accordance with embodiments of the present disclosure described above.

Fig. 9 illustrates an example apparatus 900 for graph-based enhanced collaborative filtering of content recommendations in accordance with an embodiment of the present disclosure.

The apparatus 900 may include: 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: the method includes 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 overall interest representations for a population of users based on a set of meta-interests, each meta-interest characterizing an element of interest, generating a user interest representation of the target user based on the set of historical content item representations and the set of overall interest representations, and predicting a click probability of the target user clicking on 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 also perform any other steps/processes of a method for graph-based enhanced collaborative filtering of content recommendations according to embodiments of the present disclosure described above.

Embodiments of the present disclosure propose a computer program product for graph-based enhanced collaborative filtering of content recommendation, comprising a computer program for execution by at least one processor for: generating candidate content item representations of the candidate content items; generating a set of historical content item representations corresponding to a set of historical content items of the target user; generating a set of overall interest representations for the population of users based on a set of meta-interests, each meta-interest characterizing an element of interest; generating a user interest representation of the target user based on the set of historical content item representations and the set of overall interest representations; and predicting a click probability of the target user clicking on the candidate content item based on the candidate content item representation and the user interest representation. Furthermore, the computer program may also be executed to implement any other steps/processes of the method for graph-based enhanced collaborative filtering of content recommendation according to embodiments of the present disclosure described above.

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

It should be understood that all operations in the methods described above are merely exemplary, and the present disclosure is not limited to any operations in the methods or the order of such operations, but rather should cover all other equivalent variations 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 context to be directed to a singular form.

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

The processor has 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 the overall design constraints imposed on the system. As an example, a processor, any portion of a processor, or any combination of processors presented in this disclosure may be implemented with a microprocessor, microcontroller, digital Signal Processor (DSP), field Programmable Gate Array (FPGA), programmable Logic Device (PLD), state machine, gated logic unit, discrete hardware circuits, and other suitable processing components configured to perform the various functions described in this disclosure. The functions of a processor, any portion of a processor, or any combination of processors presented in this disclosure may be implemented using software that is executed by a microprocessor, microcontroller, DSP, or other suitable platform.

Software should be construed broadly to mean instructions, instruction sets, code segments, program code, programs, subroutines, software modules, applications, software packages, routines, subroutines, objects, threads of execution, procedures, functions, and the like. The software may reside in a computer readable medium. Computer-readable media may include, for example, memory, which may be, for example, a magnetic storage device (e.g., hard disk, floppy disk, magnetic strips), optical disk, smart card, flash memory device, random Access Memory (RAM), read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically Erasable PROM (EEPROM), registers, or removable disk. Although the memory is shown separate from the processor in various aspects presented in this disclosure, the memory may also be located internal to the processor, such as a 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. Accordingly, 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 in the disclosure that are known or later come to be known to those of ordinary skill in the art are expressly incorporated herein and are intended to be encompassed by the claims.

Claims

1. A method for graph-based enhanced collaborative filtering of content recommendations, comprising:

generating candidate content item representations of the candidate content items;

generating a set of historical content item representations corresponding to a set of historical content items of the target user;

generating a set of overall interest representations for the population of users based on a set of meta-interests, each meta-interest characterizing an element of interest;

generating a user interest representation of the target user based on the set of historical content item representations and the set of overall interest representations; and

a click probability of the target user clicking on the candidate content item is predicted based on the candidate content item representation and the user interest representation.

2. The method of claim 1, wherein the generating candidate content item representations comprises:

the candidate content item representations are generated 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, from the knowledge-graph, a set of neighbor nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighbor nodes; and

generating the candidate content item representations based on at least one of: a set of neighbor node representations corresponding to the set of neighbor nodes, a set of relationship 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 set of historical content item representations comprises, for each historical content item in the set of historical content items:

a historical content item representation of the historical content item is generated based on a knowledge graph corresponding to the historical content item.

5. The method of claim 1, wherein the generating a user interest representation comprises:

The user interest representation is generated by aggregating the set of historical content item representations with the set of overall interest representations.

6. The method of claim 1, wherein the method is performed by a collaborative filtering model.

7. The method of claim 6, wherein the training of the collaborative filtering model comprises:

reducing the dimensions of the set of overall interest representations to obtain a set of reduced-dimension overall interest representations; and

applying distance-related constraints to the set of reduced-dimension global interest representations to enhance diversity of the set of reduced-dimension global interest representations.

8. The method of claim 7, wherein the reducing the dimension of the set of overall interest representations comprises:

principal component analysis is utilized to reduce the dimensions of the set of overall interest representations.

9. The method of claim 6, wherein the training of the collaborative filtering model comprises:

the collaborative filtering model is trained using a training data set comprising 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 popularity of each content item sample in the set.

10. The method of claim 6, wherein the training of the collaborative filtering model comprises:

the collaborative filtering model is trained using a content-based filtering model.

11. The method of claim 10, wherein the training the collaborative filtering model with a content-based filtering model comprises:

generating cross-system contrast prediction loss by adopting cross-system contrast learning through the collaborative filtering model and the content-based filtering model; and

the collaborative filtering model is optimized at least by minimizing the cross-system contrast prediction loss.

12. The method of claim 11, wherein the generating cross-system contrast prediction loss comprises:

for each training data of a plurality of training data included in a training data set, generating cross-system versus sub-prediction losses for the training data to obtain a plurality of cross-system versus sub-prediction losses for the plurality of training data; and

combining the plurality of cross-system contrast sub-prediction losses into the cross-system contrast prediction losses.

13. The method of claim 12, wherein the training data comprises a user sample and positive and negative content item samples associated with the user sample, and the generating cross-system versus sub-prediction losses comprises:

Generating, by 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, by 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 representation of the negative content item sample; and

the cross-system contrast sub-prediction loss is generated 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.

14. The method according to claim 13, wherein:

the generating a second positive content item sample representation comprises: generating, by the content-based filtering model, the second positive content item sample representation based on the textual description of the positive content item sample, and/or

The generating a second negative content item sample representation comprises: the second negative content item sample representation is generated based on the text description of the negative content item sample by the content-based filtering model.

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

16. An apparatus for graph-based enhanced collaborative filtering of content recommendations, comprising:

at least one processor; and

a memory storing computer-executable instructions that, when executed, cause the at least one processor to:

candidate content item representations of the candidate content items are generated,

a set of historical content item representations corresponding to a set of historical content items of the target user are generated,

generating a set of overall interest representations for the population of users based on a set of meta-interests, each meta-interest characterizing an element of interest,

generating a user interest representation of the target user based on the set of historical content item representations and the set of overall interest representations, an

17. The apparatus of claim 16, wherein the generating candidate content item representations comprises:

18. The apparatus of claim 17, wherein the generating the candidate content item representation comprises:

19. The apparatus of claim 16, wherein the generating a user interest representation comprises:

20. A computer program product for graph-based enhanced collaborative filtering of content recommendation, comprising a computer program for execution by at least one processor for: