Content recommendation based on graph-enhanced collaborative filtering

Abstract

The present disclosure presents methods, apparatuses and computer program products for content recommendation based on graph-enhanced collaborative filtering. A candidate content item representation of a candidate content item can be generated. A set of historical content item representations corresponding to a set of historical content items of a target user can be generated. A set of overall interest representations for all users can be generated based on a set of meta-interests, each meta-interest representing an interest element. A user interest representation of the target user can be generated based on the set of historical content item representations and the set of overall interest representations.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 the candidate content item can be predicted based on the candidate content item representation and the user interest representation.

Description

Background Technology

[0001] With the development of internet technology and the growth of online information, recommendation systems are playing an increasingly important role in many online services. Different recommendation systems exist based on the content being recommended, such as movie recommendation systems, book recommendation systems, music recommendation systems, and product recommendation systems. These systems typically capture user interests and predict content that the user might be interested in based on those interests, then recommend it to the user. Summary of the Invention

[0002] This invention is provided to introduce a set of concepts, which will be further described in the following detailed description. This invention is not intended to identify key or essential features of the protected subject matter, nor is it intended to limit the scope of the protected subject matter.

[0003] Embodiments of this disclosure provide methods, apparatus, and computer program products for content recommendation based on graph-enhanced collaborative filtering. Candidate content item representations can be generated. A set of historical content item representations corresponding to a set of historical content items for a target user can be generated. A set of overall interest representations for all users can be generated based on a set of meta-interests, each meta-interest representing an interest element. A user interest representation for the target user can be generated based on the set of historical content item representations and the set of overall interest representations. The click probability of the target user clicking on a candidate content item can be predicted based on the candidate content item representations and the user interest representations.

[0004] It should be noted that one or more of the above aspects include the features specifically pointed out in the following detailed description and claims. Certain illustrative features of the one or more aspects are set forth in detail in the following specification and drawings. These features merely indicate various ways in which the principles of each aspect can be implemented, and this disclosure is intended to include all such aspects and their equivalents. Attached Figure Description

[0005] The following description will take into account several aspects disclosed, which are provided to illustrate rather than limit the aspects disclosed.

[0006] Figure 1 An exemplary process for content recommendation based on graph-enhanced collaborative filtering according to embodiments of this disclosure is illustrated.

[0007] Figure 2 An exemplary process for training a collaborative filtering model according to embodiments of this disclosure is shown.

[0008] Figure 3 An exemplary process for predicting click probability using a content-based filtering model according to embodiments of this disclosure is illustrated.

[0009] Figure 4 An exemplary process for training a content-based filtering model according to embodiments of this disclosure is shown.

[0010] Figure 5 Another exemplary process for training a collaborative filtering model according to an embodiment of this disclosure is shown.

[0011] Figure 6 An exemplary process for generating cross-system comparative prediction loss according to embodiments of this disclosure is illustrated.

[0012] Figure 7 This is a flowchart of an exemplary method for content recommendation based on graph-enhanced collaborative filtering, according to embodiments of this disclosure.

[0013] Figure 8 An exemplary apparatus for content recommendation based on graph-enhanced collaborative filtering, according to embodiments of the present disclosure, is shown.

[0014] Figure 9 An exemplary apparatus for content recommendation based on graph-enhanced collaborative filtering, according to embodiments of the present disclosure, is shown. Detailed Implementation

[0015] This disclosure will now be discussed with reference to several exemplary embodiments. It should be understood that this discussion of embodiments is merely intended to enable those skilled in the art to better understand and thus implement the embodiments of this disclosure, and is not intended to teach any limitation on the scope of this disclosure.

[0016] Collaborative filtering (CF) is a widely used recommendation technique. It aims to identify similar users to the target user, determine the target user's interests based on these similar users' interests, and then recommend content to the target user based on those interests. In this paper, the target user can refer to the user to whom content recommendations are performed. Collaborative filtering can be implemented using machine learning models. In this paper, the machine learning model employing collaborative filtering-based content recommendation methods can be called a collaborative filtering model. The collaborative filtering model can predict the click probability of the target user clicking each candidate content item in a set of candidate content items, thus obtaining a set of click probabilities. In this paper, a content item can refer to an individual item with specific content. For example, a movie, a book, or a piece of music can be considered a content item. The content items to be recommended to the target user can be determined by ranking this set of click probabilities.

[0017] Collaborative filtering models can generate candidate content item representations and target user interest representations for candidate content items, and predict the click probability of a target user clicking on a candidate content item based on these representations. Typically, the target user's previously viewed, clicked, or watched historical content items can indicate their interests; therefore, the target user's interest representation can be generated based on the representation of historical content items. When generating representations of content items such as candidate and historical content items, a corresponding knowledge graph can be considered. A knowledge graph can contain a large number of nodes representing numerous entities and a large number of edges representing rich relationships between entities. In a knowledge graph, neighboring nodes adjacent to a specific node can be considered as structured information about that specific node, such as attribute information, and edges from neighboring nodes to that specific node can be considered as relationships between neighboring nodes and that specific node. Therefore, when using collaborative filtering for content recommendation, knowledge graphs can be used to enrich the candidate content item representations and user interest representations, thereby more accurately predicting the click probability of a target user clicking on a candidate content item. The method of content recommendation using knowledge graph-based collaborative filtering can be called the graph-enhanced collaborative filtering content recommendation method.

[0018] Embodiments of this disclosure propose an improved content recommendation method based on graph-augmented collaborative filtering. Candidate content item representations of candidate content items and user interest representations of target users can be generated, and the click probability of a target user clicking on a candidate content item can be predicted based on the generated candidate content item representations and user interest representations. When generating user interest representations, a set of historical content item representations corresponding to a set of historical content items for the target user and a set of overall interest representations for all users can be generated, and the user interest representation can be generated based on the set of historical content item representations and the set of overall interest representations. The overall interest representation can correspond to an overall interest. In this document, overall interest can refer to the interests of all users rather than a specific user. Embodiments of this disclosure propose that a set of overall interest representations can be generated based on a set of meta-interests. In this document, meta-interest can refer to one of the interest elements in a set of interest elements used to constitute the overall interest for all users. Each meta-interest can have a corresponding trainable embedding vector. Utilizing meta-interests to generate the overall interest representation can more broadly capture different interest points in the overall interest for all users. The above-described improved graph-augmented collaborative filtering content recommendation method can be implemented using a collaborative filtering model according to embodiments of this disclosure.

[0019] In one aspect, embodiments of this disclosure propose generating candidate content item representations based on a knowledge graph corresponding to the candidate content items. For example, a set of neighboring nodes adjacent to the candidate content item and a set of edges corresponding to that set of neighboring nodes can be identified from the knowledge graph. The candidate content item representation can be generated based on the neighboring node representations corresponding to that set of neighboring nodes, the relation representations corresponding to that set of edges, and the importance of that set of neighboring nodes to the candidate content item. The importance of neighboring nodes to the candidate content item can be used to control the information passed from neighboring nodes to the candidate content item. The more important a neighboring node is to the candidate content item, the more information it can pass to the candidate content item, thus influencing 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 can be generated based on the knowledge graph corresponding to that historical content item.

[0020] In another aspect, embodiments of this disclosure propose employing soft distance correlation to enhance the diversity of a set of overall interest representations used to generate user interest representations when training a collaborative filtering model. For example, the dimensionality of a set of overall interest representations can be reduced to obtain a set of dimensionality-reduced overall interest representations, and distance correlation constraints can be applied to this set of dimensionality-reduced overall interest representations to enhance their diversity. Applying distance correlation constraints to a set of overall interest representations ensures that the individual overall interest representations are separated from each other in the vector space, thus guaranteeing that each overall interest representation is as different as possible, allowing different overall interest representations to represent different interests from different perspectives. Furthermore, reducing the dimensionality of a set of overall interest representations first and applying distance correlation constraints only to a set of dimensionality-reduced overall interest representations ensures diversity of overall interest representations at low dimensions while maintaining flexibility at high dimensions. In this way, the different interests of users can be learned more comprehensively, resulting in the generation of more accurate user interest representations.

[0021] In another aspect, embodiments of this disclosure propose determining negative content item samples using reciprocal ratio negative sampling when constructing the training dataset for training a collaborative filtering model. Here, negative content item samples can refer to content item samples that the user is not interested in. Negative content item samples can be sampled from a set of candidate content item samples based on the popularity of each content item sample in that set. For example, content items with more user interaction can be considered popular. Popular content items may have a lower probability of being sampled as negative content item samples. Unlike existing methods that determine negative content item samples through random sampling, determining negative content item samples using reciprocal ratio negative sampling avoids identifying potentially interesting but previously unencountered content items as negative content item samples for the user, thereby improving the quality of negative content item samples.

[0022] In another aspect, embodiments of this disclosure propose using a content-based filtering (CBF) model to train a collaborative filtering model. The CBF model can utilize unstructured information about candidate and / or historical content items, such as textual descriptions of the content of the candidate and / or historical content items, when generating representations of these items. The historical content item representations can be further used to generate user interest representations for the target user. Using a CBF model to train a collaborative filtering model allows the transfer of knowledge learned by the CBF model, such as knowledge gained from the unstructured information of content items, to the collaborative filtering model. The collaborative filtering model can be trained using a CBF model through cross-system contrastive learning. This approach provides a lightweight way to fuse collaborative filtering and content-based filtering models. Collaborative filtering models trained using content-based filtering models can incorporate textual descriptions of candidate content items and / or historical content items when generating representations of candidate content items and / or user interests, thus generating richer representations of candidate content items and / or user interests. Furthermore, for certain new content items, such as newly released movies or newly published books, recommendations can be made based on their textual descriptions even if there is limited user interaction with them or they were not covered in previous training. This effectively addresses the cold start problem.

[0023] Figure 1 An exemplary process 100 for content recommendation based on graph-enhanced collaborative filtering according to embodiments of the present disclosure is illustrated. In process 100, a target user can be predicted using a collaborative filtering model 110. Click on candidate content item 102 ( The click probability is 112.

[0024] Candidate content item 102 It can come from sources that can be recommended to the target user. Content items from a set of candidate content items. Candidate content item 102 This can include, for example, movies, books, music, videos, product information, news, etc.

[0025] Candidate content items 102 can be generated using the content item encoder 120. The candidate content item representation is 122. The content item encoder 120 may, for example, include... A gated path graph convolutional network with 100 convolutional layers. Preferably, the content item encoder 120 can generate candidate content item representations 122 based on the knowledge graph 106. The knowledge graph 106 can be labeled as... ,in It is a set of nodes in this knowledge graph, and This is a set of edges between the nodes in this group. Knowledge graph 106 can be related to candidate content item 102. The corresponding knowledge graph. (Based on candidate content item 102) Taking movies as an example, knowledge graph 106 could be a knowledge graph containing a set of nodes with the entity type "movie" and a set of edges between these nodes. Candidate content item 102 This can correspond to knowledge graph 106 Nodes in It is possible to identify nodes from knowledge graph 106. A group of adjacent neighboring nodes and a set of edges corresponding to the neighboring nodes of that group. In the convolutional layers of a gated path graph convolutional network At this point, it can be based on nodes. A group of neighboring nodes In convolutional layers The representation at the location, corresponding to the node With the neighboring nodes of this group A set of edges representing a relationship between two neighbors, and the set of neighboring nodes. For nodes The importance of generating nodes, etc. exist The representation of the location is shown in the following formula:

[0026] (1)

[0027] in, From node to neighboring nodes Relationship Relationship representation, From neighboring nodes To the node A gating function that controls the information transmitted; it can be used to control information transmitted from neighboring nodes. Passed to the node The information is weighted. It can be a node or candidate content items The initial representation can be a randomly initialized representation or a representation obtained through known knowledge graph embedding vector methods. Similarly, relation representations... It can be a randomly initialized representation or a representation obtained through a known knowledge graph embedding vector method.

[0028] Gating function It can reflect neighboring nodes For nodes The importance of gating functions can be calculated using the following formula. :

[0029] (2)

[0030] in, This is the sigmoid function used to restrict gated values ​​to between 0 and 1. For nodes... More important neighboring nodes can provide assistance to the node. The more information transmitted, the greater the influence on the nodes. The representation of candidate content items The representation of . It should be understood that although in formula (1) the candidate content item representation is generated based on three factors including neighbor node representation, relation representation and the importance of neighbor nodes to the candidate content item, in some embodiments it is also possible to generate the candidate content item representation based on only one or two of these three factors.

[0031] To overcome the oversmoothing problem of graph convolution, node aggregation can be used. Obtaining nodes by representation at all intermediate layers The final representation, i.e., the candidate content item representation 122 As shown in the following formula:

[0032] (3)

[0033] In order to generate target users User interest representation 142 can be used to obtain target users. A set of historical content items 104. This set of historical content items 104 may include the target user. Multiple previously viewed, clicked, or viewed historical content items, such as historical content item 104-1 to historical content item 104- ,in This refers to the number of historical content items. Historical content items can include, for example, movies, books, music, videos, product information, news, etc. Target users A set of historical content items 104 can indicate the target user User interests. Taking movies as an example of historical content, the target users... Previously watched movies can indicate which movies a user is interested in.

[0034] A set of historical content item representations 132 corresponding to a set of historical content items 104 can be generated using a set of content item encoders 130. The set of content item encoders 130 may include, for example, content item encoders 130-1 to content item encoders 130- A set of historical content item representations 132 may include, for example, historical content item representations 132-1 to historical content item representations 132- Each content item encoder in a set of content item encoders 130 may have a similar structure to content item encoder 120, for example, it may include... A gated path graph convolutional network with one convolutional layer. Historical content item representations 132-1 to 132- can be generated through a process similar to that used to generate candidate content item representation 122. For example, for each historical content item in a set of historical content items 104, a historical content item representation can be generated based on the knowledge graph corresponding to that historical content item, such as knowledge graph 106.

[0035] Subsequently, the target user can be generated using user encoder 140. User interest representation 142. For example, it can be based on a set of overall interests for all users 108 This generates a set of overall interest representations for all users. It can include interactions between all users and content items. Target users can be generated based on a set of historical content item representations (132) and a set of overall interest representations. User interest is represented by 142.

[0036] According to embodiments of this disclosure, the overall interest group Each overall interest It can be composed of a set of meta-interests Composition. Each meta-interest An interest element can be represented. For each overall interest... This can be achieved through the interest of the group. To generate this overall interest Overall interest representation As shown in the following formula:

[0037] (4)

[0038] in, Meta-interest The trainable embedding vectors, and { } is the general interest as shown in the following formula. From trainable weights { The linear weights derived in}

[0039] (5)

[0040] It can be tailored to each overall interest To perform the above operations to generate a set of overall interest representations Next, this can be achieved by utilizing a set of overall interest representations. For convolutional layers A set of historical content items (represented by 132) are aggregated to generate user interests in the convolutional layer. The representation of the location is shown in the following formula:

[0041] (6)

[0042] in, User Historical interactions, It is a historical content item. In convolutional layers The indication of the location, and The weights for overall interest are obtained through an attention mechanism, as shown in the following formula:

[0043] (7)

[0044] Similar to generating candidate content item representation 122, to overcome the oversmoothing problem of graph convolution, the final representation of user interest, i.e., user interest representation 142, can be obtained by aggregating the representations of user interest at all intermediate layers. As shown in the following formula:

[0045] (8)

[0046] After generating candidate content item representation 122 User interest representation 142 Then, the target user can be predicted using prediction layer 150. Click on candidate content item 102 The click probability is 112. Click probability can be tagged as... In one implementation, this can be achieved by representing user interests 142. And candidate content items represent 122 Applying dot product operations to predict click probability As shown in the following formula:

[0047] (9)

[0048] Since the collaborative filtering model 110 can utilize knowledge graphs to generate candidate content item representations and / or historical content item representations, and can generate user interest representations based on a set of meta-interests, the collaborative filtering model 110 can also be called a knowledge-graph-enhanced meta-interest network.

[0049] It should be understood that the above text, in combination with... Figure 1 The described process for content recommendation based on graph-augmented collaborative filtering is merely exemplary. Depending on the specific application requirements, the steps in the process can be replaced or modified in any way, and the process may include more or fewer steps. For example, although in process 100, a knowledge graph corresponding to the candidate content item and / or historical content item is considered when generating the candidate content item representation and / or historical content item representation, in some embodiments, it is feasible not to consider the knowledge graph. In this case, the corresponding candidate content item representation and / or historical content item representation can be generated solely based on the candidate content item and / or historical content item itself. Furthermore, the specific order or hierarchy of the steps in process 100 is merely exemplary, and the process for content recommendation based on graph-augmented collaborative filtering can be performed in an order different from the described order.

[0050] Collaborative filtering models can be trained in various ways, such as... Figure 1 Collaborative filtering model 110. Figure 2 An exemplary process 200 for training a collaborative filtering model according to an embodiment of this disclosure is illustrated. When the collaborative filtering model trained by process 200 is actually deployed, it can predict the click probability of a target user clicking on a candidate content item.

[0051] Such as combination Figure 1 As described, when generating user interest representations for a target user, a set of overall interest representations for all users is considered. Embodiments of this disclosure propose employing soft distance correlation to enhance the diversity of this set of overall interest representations.

[0052] For example, at position 202, the dimensionality of a set of overall interest representations can be reduced to obtain a dimensionality-reduced set of overall interest representations. In one implementation, Principal Component Analysis (PCA) can be used to reduce the dimensionality of a set of overall interest representations, as shown in the following formula:

[0053] (10)

[0054] in, This is the ratio of principal components to be retained after PCA. This ratio can be a value between 0 and 1.

[0055] At position 204, a soft distance correlation prediction loss can be generated by applying a distance correlation constraint to the dimensionality-reduced overall interest representation. As shown in the following formula:

[0056] (11)

[0057] in, Used to calculate distance covariance, and Used to calculate distance variance. It should be understood that, in formula (10), hour, The result is the original distance-related prediction loss.

[0058] Applying distance-related constraints to a set of overall interest representations ensures that these representations are separated from each other in the vector space, guaranteeing that each representation is as distinct as possible. This allows different overall interest representations to characterize different interests from various perspectives. Furthermore, reducing the dimensionality of a set of overall interest representations and applying distance-related constraints only to this reduced set ensures diversity in the lower dimensions while maintaining flexibility in the higher dimensions. In this way, a more comprehensive understanding of the user's diverse interests can be achieved, leading to more accurate user interest representations.

[0059] Alternatively, a collaborative filtering model can be trained using a training dataset that includes multiple positive content item samples and multiple negative content item samples.

[0060] For example, at position 210, multiple positive content item samples can be obtained. For example, users from a set of candidate content item samples can be... Previously viewed, clicked, or viewed content items are considered to be relevant to the user. Sample of positive content items .

[0061] At position 212, multiple negative content item samples can be obtained. Embodiments of this disclosure propose determining negative content item samples by inverse ratio negative sampling. For example, multiple negative content item samples can be obtained by sampling from a set of candidate content item samples based on the popularity of each content item sample in that set. For example, content items with high user interaction can be considered popular. Popular content items have a lower probability of being sampled as negative content items. For content item samples It was sampled as a negative content item sample. The probability can be expressed by the following formula:

[0062] (12)

[0063] in, Represents a sample of all users and content items. Interaction counting. Unlike existing methods that determine negative content items through random sampling, determining negative content items through inverse ratio negative sampling avoids identifying potentially interesting but previously unseen content items as negative content items for the user, thus improving the quality of negative content item samples.

[0064] At position 214, it can be based on multiple positive content item samples. and multiple negative content item samples To construct the training dataset .

[0065] At point 216, paired Bayesian Personalized Ranking (BPR) can be generated based on the constructed dataset to predict the loss. The BPR prediction loss aims to assign higher scores to content items viewed by the user compared to those not of interest, using the concept of contrastive learning. The BPR prediction loss can be generated, for example, using the following formula. :

[0066] (13)

[0067] Alternatively or additionally, at point 218, an L2 regularization prediction loss can be generated based on the constructed dataset. ,in ,and It is the L2 norm of the user interest representation / content item representation.

[0068] Subsequently, at position 220, the comprehensive prediction loss of the collaborative filtering model can be generated based on at least one of the distance-related prediction loss, BPR prediction loss, and L2 regularized prediction loss. As shown in the following formula:

[0069] (14)

[0070] in, and It is a hyperparameter used to control the weights of the prediction loss.

[0071] At point 222, the loss can be predicted by making the comprehensive forecasting method work. Minimize the training parameters to train the collaborative filtering model.

[0072] It should be understood that the above text, in combination with... Figure 2 The described process for training a collaborative filtering model is merely exemplary. Depending on the specific application requirements, the steps in the process can be replaced or modified in any way, and the process may include more or fewer steps. For example, although in process 200, distance correlation prediction loss, BPR prediction loss, and 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 on only one or two of these three factors. Furthermore, the specific order or hierarchy of the steps in process 200 is merely exemplary, and the process for training a collaborative filtering model may be performed in an order different from that described.

[0073] The above text combined Figure 1This paper describes a collaborative filtering model used to generate candidate content item representations and target user interest representations, and then predicts the click probability of the target user clicking on a candidate content item based on the generated candidate content item representations and user interest representations. The collaborative filtering model considers structured information of the content items, such as various attribute information from knowledge graphs, when generating candidate content item representations and / or user interest representations. Embodiments of this disclosure propose using a content-based filtering model to train the collaborative filtering model. Similar to the collaborative filtering model, the content-based filtering model can generate candidate content item representations and target user interest representations, and predict the click probability of the target user clicking on the candidate content item based on these representations. Furthermore, the user interest representation can also be generated based on the user's historical content item representations. However, the content-based filtering model can utilize unstructured information of the candidate content items and / or historical content items, such as text descriptions reflecting the content of the candidate content items and / or historical content items, when generating representations of candidate content items and / or historical content items. Cross-system comparative learning allows for the training of collaborative filtering models using content-based filtering models, providing a lightweight approach to fusing collaborative filtering and content-based filtering. Collaborative filtering models trained on content-based models can incorporate textual descriptions of candidate content items and / or historical content items when generating candidate content item representations and / or user interest representations, resulting in richer representations. Furthermore, for new content items, such as newly released movies or newly published books, recommendations can be made based on their textual descriptions, even if there is limited user interaction or they were not covered in previous training, effectively addressing the cold start problem.

[0074] Figure 3 An exemplary process 300 for predicting click probability using a content-based filtering model according to an embodiment of this disclosure is illustrated. In process 300, a target user can be predicted using a content-based filtering model 310. Click on candidate content item 302 The click probability is 312.

[0075] Candidate content item 302 It can come from sources that can be recommended to the target user. Content item 302 is a candidate content item from a set of candidate content items. This can include, for example, movies, books, music, videos, product information, news, etc. Candidate content item 302 can be obtained. Text description 306 The text description 306 can be obtained through various known methods. In one implementation, it can be selected from candidate content item 302. 306 text descriptions were extracted from the corresponding knowledge graph. In another implementation, a trained machine learning model can be used to search the internet for candidate content item 302. Related articles, and identified important paragraphs from the searched articles as candidate content items 302 Text description 306 .

[0076] It can be achieved through the content item encoder 320, based on the text description 306 To generate candidate content item 302 The candidate content items represent 322 The content item encoder 320 can be, for example, 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 a BERT model as an example, the candidate content item representation 322... It can be represented by the following formula:

[0077] (15)

[0078] It is possible to acquire target users A set of historical content items 304. This set of historical content items 304 may include the target user. Multiple previously viewed, clicked, or viewed historical content items, such as historical content item 304-1 to historical content item 304- ,in This refers to the number of historical content items. Historical content items can include, for example, movies, books, music, videos, product information, news, etc. Target users A set of historical content items with a 304 error can indicate the target user. User interests. A set of text descriptions 308 corresponding to a set of historical content items 304 can be obtained. For example, the text description 308-1 of historical content item 304-1 can be obtained, and the text description 308-1 of historical content item 304- can be obtained. Text description 308- ,etc.

[0079] A set of historical content item representations 332 corresponding to a set of historical content items 304 can be generated using a set of content item encoders 330. The set of content item encoders 330 may include, for example, content item encoders 330-1 to content item encoders 330- A set of historical content items, represented as 332, may include, for example, historical content item representations 332-1 to 332-. Each content item encoder in a set of content item encoders 330 can have a similar structure to content item encoder 320, for example, it can be a pre-trained transformer-based model such as BERT or GPT-2. Historical content item representations 332-1 to historical content item representations 332- can be generated through a process similar to that used to generate candidate content item representations 322. For example, for each historical content item in a set of historical content items 304, a historical content item representation can be generated based on the text description of that historical content item.

[0080] Subsequently, the target user can be generated using user encoder 340. User interest representation 342 The user encoder 340 can be based on the target user. A set of historical content items (332) is used to generate a user interest representation (342). In one implementation, the set of historical content items represented by 332 can be used to represent the included items. 342 historical content items are weighted and summed to generate a user interest representation. As shown in the following formula:

[0081] (16)

[0082] in, It is assigned to historical content items The attention weights are obtained by passing features through two linear layers, as shown in the following formula:

[0083] (17)

[0084] (18)

[0085] in, , , and These are the weights and biases of the two fully connected layers, respectively.

[0086] After generating candidate content item representation 322 User interest representation 342 Then, the target user can be predicted using prediction layer 350. The probability of clicking candidate item 302 is 312. Click probability can be denoted as... In one implementation, this can be achieved by representing user interests 342. And candidate content items represent 322 Using dot product operations to predict click probability As shown in the following formula:

[0087] (19)

[0088] Because the content-based filtering model 310 employs a multi-head self-attention mechanism, it can also be referred to as a Neural Recommendation with Multi-Head Self-Attention (NRMS) model. In particular, when the content-based filtering model 310 uses a pre-trained BERT model as the content item encoder, it can also be called an NRMS-BERT model.

[0089] It should be understood that the above text, in combination with... Figure 3 The described process of using a content-based filtering model to predict click probabilities is merely exemplary. Depending on the specific application requirements, the steps in this process can be replaced or modified in any way, and the process may include more or fewer steps. Furthermore, the specific order or hierarchy of the steps in process 300 is only exemplary, and the process of using a content-based filtering model to predict click probabilities can be performed in an order different from the described order.

[0090] According to embodiments of this disclosure, a content-based filtering model can be used to train a collaborative filtering model. For example, a content-based filtering model can be used... Figure 3 The content-based filtering model 310 is used for training. Figure 1 The collaborative filtering model 110 is described above. Preferably, the content-based filtering model can be pre-trained before using the content-based filtering model to train the collaborative filtering model. A negative sampling method can be used to train the content-based filtering model. Figure 4 An exemplary process 400 for training a content-based filtering model according to an embodiment of this disclosure is illustrated. The content-based filtering model trained by process 400, when actually deployed, can predict the click probability of a target user clicking on candidate content items.

[0091] First, a training dataset can be constructed for training the content-based filtering model. In one implementation, a list-wise strategy can be used to construct the training dataset. For example, at position 410, multiple positive content item samples can be obtained. For instance, content items that the user has previously viewed, clicked, or watched from a set of candidate content item samples can be considered as positive content item samples.

[0092] At position 420, multiple sets of negative content item samples corresponding to multiple positive content item samples can be obtained. For example, for each positive content item sample, the set of content item samples that are presented in the same session as that positive content item sample but have not been clicked by the user can be regarded as the set of negative content item samples corresponding to that positive content item sample.

[0093] At position 430, a training dataset for training a content-based filtering model can be constructed based on multiple positive content item samples and a set of multiple negative content item samples corresponding to the multiple positive content item samples.

[0094] Subsequently, multiple posterior click probabilities corresponding to multiple positive content item samples can be generated. For example, at position 440, the click probability of the positive content item sample corresponding to each positive content item sample can be predicted. The click probability of a positive content item sample corresponding to a positive content item sample can be denoted as: .

[0095] At position 450, for each negative content item sample in the set of negative content item samples corresponding to the positive content item sample, the click probability of the negative content item sample corresponding to that negative content item sample can be predicted to obtain the set of click probabilities of negative content item samples corresponding to that set of negative content item samples. (This is related to the previous sentence, which is incomplete and requires further context.) The set of click probabilities for negative content items corresponding to the set of negative content item samples of positive content item samples can be labeled as follows: ,in This represents the number of negative content item samples included in the set of negative content item click probabilities. In this way, the click probability prediction problem can be expressed as a pseudo-... +1 category task.

[0096] At position 460, the posterior click probability corresponding to the positive content item sample can be calculated based on the set of click probabilities for positive and negative content item samples. (This is in contrast to the previous point.) The posterior click probability corresponding to each positive content item sample can be labeled as: In one implementation, the click probability of positive content item samples can be calculated using the softmax function. and the set of click probabilities for negative content items Normalization is performed to calculate the posterior click probability corresponding to the positive content item sample, as shown in the following formula:

[0097] (20)

[0098] The operations described in steps 440 to 460 can be performed on each of the multiple positive content item samples in the training dataset, so that at 470, multiple posterior click probabilities corresponding to the multiple positive content item samples can be obtained.

[0099] At position 480, the prediction loss can be generated based on multiple posterior click probabilities. In one implementation, the prediction loss can be generated by calculating the negative log-likelihood of multiple posterior click probabilities, as shown in the following formula:

[0100] (twenty one)

[0101] in, It is a collection of positive content item samples consisting of multiple positive content item samples.

[0102] At 490, the content-based filtering model can be optimized by minimizing the prediction loss.

[0103] It should be understood that the above text, in combination with... Figure 4 The described process for training a content-based filtering model is merely exemplary. Depending on the specific application requirements, the steps in the process for training the content-based filtering model can be replaced or modified in any way, and the process may include more or fewer steps. For example, although a list-based strategy is used to construct the training dataset in process 400, other methods can also be used to construct the training dataset, such as combining [methods] when determining negative content item samples. Figure 2 The described inverse ratio negative sampling method. Furthermore, the specific order or hierarchy of the steps in process 400 is merely exemplary, and the process for training the content-based filtering model may be performed in a different order than described.

[0104] Embodiments of this disclosure propose training a collaborative filtering model using a content-based filtering model through cross-system comparative learning. Figure 5 Another exemplary process 500 for training a collaborative filtering model according to an embodiment of this disclosure is shown. Compared to combining Figure 2In process 200, in process 500, in addition to generating the comprehensive prediction loss of the collaborative filtering model based on at least one of soft distance correlation prediction loss, BPR prediction loss and L2 regularization prediction loss, cross-system comparative prediction loss can also be generated by cross-system comparative learning through collaborative filtering model and content-based filtering model.

[0105] Steps 502 to 520 can respectively correspond to Figure 2 Steps 202 to 220 in the process.

[0106] At point 530, cross-system contrastive prediction loss can be generated using cross-system contrastive learning through collaborative filtering models and content-based filtering models. The following will be combined with Figure 6 To illustrate the generation of cross-system comparative prediction loss An exemplary process.

[0107] At point 540, loss can be predicted based on comprehensive analysis. Comparison with cross-system prediction loss To generate a cross-system enhanced comprehensive prediction loss for collaborative filtering models As shown in the following formula:

[0108] (twenty two)

[0109] in, Used to control cross-system comparison prediction loss The weight.

[0110] At 550, the comprehensive prediction loss can be enhanced by improving the cross-system capabilities. Minimize to optimize the collaborative filtering model.

[0111] It should be understood that the above text, in combination with... Figure 5 The described process for training a collaborative filtering model using a content-based filtering model is merely exemplary. Depending on the specific application requirements, the steps in this process can be replaced or modified in any way, and the process may include more or fewer steps. For example, although in process 500 the collaborative filtering model is trained based on both the combined prediction loss and the cross-system contrastive prediction loss, in some embodiments, the collaborative filtering model may be trained based solely on the cross-system contrastive prediction loss. In this case, the collaborative filtering model can be optimized by minimizing the cross-system contrastive prediction loss. Furthermore, the specific order or hierarchy of the steps in process 500 is merely exemplary, and the process for training a collaborative filtering model using a content-based filtering model may be performed in a different order than described.

[0112] Figure 6 An exemplary process 600 for generating cross-system contrastive prediction loss according to an embodiment of this disclosure is shown. Process 600 may correspond to Figure 5 Step 530 in the process.

[0113] A cross-system contrastive prediction loss can be generated using a pre-built training dataset 610. The training dataset 610 can be obtained, for example, through... Figure 5 Steps 510 to 514 in the process, namely Figure 2 Steps 210 to 214 in the process construct the training data set. For the training dataset 610 620 training data points It can generate 620 training data. Cross-system comparison sub-prediction loss 650 .

[0114] Training data 620 This may include, for example, user sample 622 624 positive content items Sample 626 negative content items Sample 624 positive content items Sample 626 negative content items Can be used with user sample 622 Related. For example, sample 624 of the positive content items. It could be user sample 622 Viewed content items, while negative content items sample 626. It could be user sample 622 Content items you are not interested in.

[0115] Collaborative filtering model 630 can, for example, correspond to Figure 1 Collaborative filtering model 110 is used. User samples 622 can be generated using collaborative filtering model 630. The first user sample interest representation 632 624 positive content items The first positive content item sample represents 634 Sample 626 negative content items The first negative content item sample represents 636 .

[0116] Content-based filtering model 640 can, for example, correspond to Figure 3 The content-based filtering model 310 in the text. User samples 622 can be generated using the content-based filtering model 640. The second user sample interest representation 642 624 positive content items The second positive content item sample represents 644 Sample 626 negative content items The second negative content item sample represents 646 Content-based filtering model 640 can be based on positive content item samples 624. The text description is used to generate the second positive content item sample representation 644. Similarly, the content-based filtering model 640 can be based on negative content item samples 626. The text description is used to generate the second negative content item sample representation 646. .

[0117] Subsequently, based on the interest representation of the first user sample 632 The first positive content item sample represents 634. The first negative content item sample represents 636. Second user sample interest representation 642 The second positive content item sample represents 644. The second negative content item sample represents 646 To generate training data 620 Cross-system comparison sub-prediction loss 650 For example, cross-system contrastive learning can be used to generate a cross-system contrastive sub-prediction loss of 650. As shown in the following formula:

[0118] (twenty three)

[0119] It can be used for training datasets Perform the above operation on each of the multiple training data sets to obtain multiple cross-system contrastive sub-prediction losses. These multiple cross-system comparisons predict the loss. This can be combined into a cross-system contrastive prediction loss 660 for collaborative filtering model 630. As shown in the following formula:

[0120] (twenty four)

[0121] Cross-system comparative prediction loss This can guide collaborative filtering models to incorporate content-sensitive information from content-based filtering models, such as information related to text descriptions. For example, it can be used to compare prediction losses across systems. This can guide collaborative filtering models to generate content item representations that interact not only with the user interest representations output by the collaborative filtering model itself but also with the user interest representations output by content-based filtering models. Similarly, cross-system comparison of prediction losses... This approach can guide collaborative filtering models to generate user interest representations that interact not only with the content item representations output by the collaborative filtering model itself but also with the content item representations output by content-based filtering models. This method allows the collaborative filtering model to learn the textual descriptions of content items output by the content-based filtering model when generating content item representations and / or user interest representations. Therefore, the knowledge learned by the content-based filtering model can be transferred to the collaborative filtering model.

[0122] It should be understood that the above text, in combination with... Figure 6 The described process for generating cross-system contrastive prediction loss is merely exemplary. Depending on the specific application requirements, the steps in the process for generating cross-system contrastive prediction loss can be replaced or modified in any way, and the process may include more or fewer steps. Furthermore, the specific order or hierarchy of the steps in process 600 is merely exemplary, and the process for generating cross-system contrastive prediction loss may be performed in an order different from that described.

[0123] Content recommendation based on graph-enhanced collaborative filtering according to embodiments of this disclosure can be applied to various types of recommendations, such as movie recommendations, book recommendations, music recommendations, video recommendations, product recommendations, and news recommendations. In particular, when using a collaborative filtering model according to embodiments of this disclosure, such as... Figure 1 The collaborative filtering model 110 shown, when deployed online, is particularly suitable for recommendations in relatively static domains with slow content updates or high user interaction, such as movie recommendations, book recommendations, music recommendations, video recommendations, and product recommendations, because it fully considers user interaction. When using content-based filtering models according to embodiments of this disclosure, such as... Figure 3 The content-based filtering model 310 shown is particularly suitable for recommendations in relatively dynamic fields with fast content updates or less user interaction, such as news recommendations, because it can make recommendations based on the text information of content items.

[0124] Figure 7 This is a flowchart of an exemplary method 700 for content recommendation based on graph-enhanced collaborative filtering according to embodiments of this disclosure.

[0125] At position 710, a candidate content item representation of the candidate content item can be generated.

[0126] At position 720, a set of historical content item representations corresponding to a set of historical content items for the target user can be generated.

[0127] At point 730, a set of overall interest representations for all users can be generated based on a set of meta-interests, where each meta-interest represents an interest element.

[0128] At point 740, the user interest representation of the target user can be generated based on the set of historical content item representations and the set of overall interest representations.

[0129] At 750, the click probability of the target user clicking the candidate content item can be predicted based on the candidate content item representation and the user interest representation.

[0130] In one implementation, generating the candidate content item representation may include: generating the candidate content item representation based on a knowledge graph corresponding to the candidate content item.

[0131] The generation of the candidate content item representation may include: identifying a set of neighboring nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighboring nodes from the knowledge graph; and generating the candidate content item representation based on at least one of the following: a set of neighboring node representations corresponding to the set of neighboring nodes, a set of relation representations corresponding to the set of edges, and the importance of the set of neighboring nodes to the candidate content item.

[0132] In one implementation, generating a set of historical content item representations may include, 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 the knowledge graph corresponding to the historical content item.

[0133] In one implementation, generating the user interest representation may include: generating the user interest representation by aggregating the set of historical content item representations using the set of overall interest representations.

[0134] In one implementation, method 700 may be performed via a collaborative filtering model.

[0135] Training the collaborative filtering model may include: reducing the dimensionality of the set of overall interest representations to obtain a set of dimensionality-reduced overall interest representations; and applying distance-related constraints to the set of dimensionality-reduced overall interest representations to enhance the diversity of the set of dimensionality-reduced overall interest representations.

[0136] The reduction of the dimensionality of the set of overall interest representations may include: using principal component analysis to reduce the dimensionality of the set of overall interest representations.

[0137] Training the collaborative filtering model may include: training the collaborative filtering model using a training dataset comprising multiple positive content item samples and multiple negative content item samples, wherein the multiple negative content item samples are obtained by sampling from the set of candidate content item samples based on the popularity of each content item sample in the set of candidate content item samples.

[0138] Training the collaborative filtering model may include using a content-based filtering model to train the collaborative filtering model.

[0139] The step of training the collaborative filtering model using a content-based filtering model may include: generating a cross-system contrastive prediction loss by employing cross-system contrastive learning using the collaborative filtering model and the content-based filtering model; and optimizing the collaborative filtering model by at least minimizing the cross-system contrastive prediction loss.

[0140] The generation of cross-system contrastive prediction loss may include: generating a cross-system contrastive sub-prediction loss for each training data in the training data set, 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.

[0141] The training data may include user samples and positive and negative content item samples associated with the user samples. Generating the cross-system contrastive sub-prediction loss may include: generating a first user sample interest representation, a first positive content item sample representation, and a first negative content item sample representation for the user samples using the collaborative filtering model; generating a second user sample interest representation, a second positive content item sample representation, and a second negative content item sample representation for the user samples using the content-based filtering model; 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.

[0142] The generation of the second positive content item sample representation may include: generating the second positive content item sample representation based on the text description of the positive content item sample using the content-based filtering model. The generation of the second negative content item sample representation may include: generating the second negative content item sample representation based on the text description of the negative content item sample using the content-based filtering model.

[0143] The candidate content items or the historical content items may include at least one of movies, books, music, videos, product information, and news.

[0144] It should be understood that method 700 may also include any steps / processes for content recommendation based on graph-enhanced collaborative filtering according to the embodiments of this disclosure described above.

[0145] Figure 8 An exemplary apparatus 800 for content recommendation based on graph-enhanced collaborative filtering according to embodiments of the present disclosure is shown.

[0146] 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; an overall interest representation generation module 830, for generating a set of overall interest representations for all users based on a set of meta-interests, each meta-interest representing an interest element; a user interest representation generation module 840, for generating a user interest representation for 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 the click probability of the target user clicking the candidate content item based on the candidate content item representations and the user interest representations. Furthermore, the apparatus 800 may also include any other modules configured for content recommendation based on graph-enhanced collaborative filtering according to the embodiments of the present disclosure described above.

[0147] Figure 9 An exemplary apparatus 900 for content recommendation based on graph-enhanced collaborative filtering according to an embodiment of this disclosure is shown.

[0148] The apparatus 900 may include: at least one processor 910; and a memory 920 storing computer-executable instructions. When executed, the computer-executable instructions cause the at least one processor 910 to: generate candidate content item representations of candidate content items; generate a set of historical content item representations corresponding to a set of historical content items for a target user; generate a set of overall interest representations for all users based on a set of meta-interests, each meta-interest representing an interest element; generate a user interest representation for the target user based on the set of historical content item representations and the set of overall interest representations; and predict the click probability of the target user clicking on the candidate content items based on the candidate content item representations and the user interest representations.

[0149] In one implementation, generating the candidate content item representation may include: generating the candidate content item representation based on a knowledge graph corresponding to the candidate content item.

[0150] The generation of the candidate content item representation may include: identifying a set of neighboring nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighboring nodes from the knowledge graph; and generating the candidate content item representation based on at least one of the following: a set of neighboring node representations corresponding to the set of neighboring nodes, a set of relation representations corresponding to the set of edges, and the importance of the set of neighboring nodes to the candidate content item.

[0151] In one implementation, generating the user interest representation may include: generating the user interest representation by aggregating the set of historical content item representations using the set of overall interest representations.

[0152] It should be understood that the processor 910 may also perform any other steps / processes of the method for content recommendation based on graph-enhanced collaborative filtering according to the above embodiments of this disclosure.

[0153] Embodiments of this disclosure propose a computer program product for content recommendation based on graph-augmented collaborative filtering, comprising a computer program executed by at least one processor for: generating candidate content item representations of candidate content items; generating a set of historical content item representations corresponding to a set of historical content items for a target user; generating a set of overall interest representations for all users based on a set of meta-interests, each meta-interest representing an interest element; generating a user interest representation for the target user based on the set of historical content item representations and the set of overall interest representations; and predicting the 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 computer program may also be executed to implement any other steps / processes for implementing the method for content recommendation based on graph-augmented collaborative filtering according to the embodiments of this disclosure described above.

[0154] Embodiments of this 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 operation of the method for content recommendation based on graph-enhanced collaborative filtering according to embodiments of this disclosure as described above.

[0155] It should be understood that all operations in the methods described above are merely exemplary, and this disclosure is not limited to any operation in the methods or the order of such operations, but should cover all other equivalent variations under the same or similar concept. Furthermore, unless otherwise specified or clearly indicated from the context as relating to the singular form, the articles “a” and “an” as used in this specification and the appended claims should generally be interpreted as meaning “one” or “one or more”.

[0156] It should also be understood that all modules in the apparatus described above can be implemented in various ways. These modules can be implemented as hardware, software, or a combination thereof. Furthermore, any of these modules can be further functionally divided into sub-modules or combined together.

[0157] Processors have been described in conjunction with various devices and methods. These processors can be implemented using electronic hardware, computer software, or any combination thereof. Whether these processors are implemented as hardware or software will depend on the specific application and the overall design constraints imposed on the system. As an example, the processors, any portions of processors, or any combination of processors given in this disclosure can be implemented using microprocessors, microcontrollers, digital signal processors (DSPs), field-programmable gate arrays (FPGAs), programmable logic devices (PLDs), state machines, gated logic units, discrete hardware circuits, and other suitable processing components configured to perform the various functions described in this disclosure. The functionality of the processors, any portions of processors, or any combination of processors given in this disclosure can be implemented using software executed by a microprocessor, microcontroller, DSP, or other suitable platform.

[0158] Software should be broadly considered to refer to instructions, instruction sets, code, code segments, program code, programs, subroutines, software modules, applications, software applications, software packages, routines, subroutines, objects, running threads, procedures, functions, etc. Software may reside on a computer-readable medium. Computer-readable media may include, for example, memory, which may be, for example, magnetic storage devices (e.g., hard disks, floppy disks, magnetic stripes), optical disks, smart cards, flash memory devices, random access memory (RAM), read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), registers, or removable disks. Although memory is shown as separate from the processor in several aspects set forth in this disclosure, memory may also reside within the processor, such as in caches or registers.

[0159] The above 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 apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects. Therefore, the claims are not intended to be limited to the aspects shown herein. All structural and functional equivalents of the elements of the various aspects described herein, known or likely to be known by one of ordinary skill in the art, are expressly incorporated herein and covered by the claims.

Claims

1. A method for content recommendation based on graph-enhanced collaborative filtering, comprising: Generate a candidate content item representation for the candidate content item; Generate a set of historical content item representations corresponding to a set of historical content items for the target user; A set of overall interest representations for all users is generated based on a set of meta-interests, where each meta-interest represents an interest element. The user interest representation of the target user is generated based on the set of historical content item representations and the set of overall interest representations; as well as The 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. The method is implemented using a collaborative filtering model, and the training of the collaborative filtering model includes: Using the collaborative filtering model and the content-based filtering model, cross-system contrastive learning is employed to generate a cross-system contrastive prediction loss; and The collaborative filtering model is optimized at least by minimizing the cross-system contrastive prediction loss.

2. The method according to claim 1, wherein, The generated candidate content items include: The candidate content item representation is generated based on the knowledge graph corresponding to the candidate content item.

3. The method according to claim 2, wherein, The generation of the candidate content item includes: Identify a set of neighboring nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighboring nodes from the knowledge graph; and The candidate content item representation is generated based on at least one of the following: 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 the importance of the set of neighbor nodes to the candidate content item.

4. The method according to claim 1, wherein, Generating a set of historical content items includes, for each historical content item in the set of historical content items: The historical content item representation of the historical content item is generated based on the knowledge graph corresponding to the historical content item.

5. The method according to claim 1, wherein, Generating user interest representations includes: The user interest representation is generated by aggregating the set of historical content item representations using the set of overall interest representations.

6. The method according to claim 1, wherein, The training of the collaborative filtering model includes: Reduce the dimensionality of the set of overall interest representations to obtain a set of dimensionality-reduced overall interest representations; and Distance correlation constraints are applied to the set of dimensionality-reduced overall interest representations to enhance the diversity of the set of dimensionality-reduced overall interest representations.

7. The method according to claim 6, wherein, The reduction of the dimensions of the overall interest representation includes: Principal component analysis is used to reduce the dimensionality of the aforementioned set of overall interest representations.

8. The method according to claim 1, wherein, The training of the collaborative filtering model includes: The collaborative filtering model is trained using a training dataset that includes multiple positive content item samples and multiple negative content item samples, wherein the multiple negative content item samples are obtained by sampling from the set of candidate content item samples based on the popularity of each content item sample in the set of candidate content item samples.

9. The method according to claim 1, wherein, The cross-system comparative prediction loss includes: For each training data point in the training dataset, a cross-system contrastive sub-prediction loss is generated for that training data, to obtain multiple cross-system contrastive sub-prediction losses for the multiple training data; and The multiple cross-system contrastive sub-prediction losses are combined into the cross-system contrastive prediction loss.

10. The method according to claim 9, wherein, The training data includes user samples and positive and negative content item samples associated with the user samples, and the cross-system contrastive sub-prediction loss includes: The collaborative filtering model generates 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. Using 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 are generated; and The cross-system contrastive 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.

11. The method of claim 10, wherein: Generating a second positive content item sample representation includes: generating the second positive content item sample representation based on the text description of the positive content item sample using the content-based filtering model, and / or Generating a second negative content item sample representation includes: generating the second negative content item sample representation based on the text description of the negative content item sample using the content-based filtering model.

12. The method according to claim 1, wherein, The candidate content items or the historical content items include at least one of movies, books, music, videos, product information, and news.

13. An apparatus for content recommendation based on graph-enhanced collaborative filtering, comprising: At least one processor; as well as A memory storing computer-executable instructions, which, when executed, cause the at least one processor to: The candidate content item representation for generating candidate content items. Generate a set of historical content item representations corresponding to a set of historical content items for the target user. A set of overall interest representations for all users is generated based on a set of meta-interests, where each meta-interest represents an interest element. Based on the set of historical content item representations and the set of overall interest representations, a user interest representation for the target user is generated, and The 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. The computer-executable instructions are executed through a collaborative filtering model, and the training of the collaborative filtering model includes: Using the collaborative filtering model and the content-based filtering model, cross-system contrastive learning is employed to generate a cross-system contrastive prediction loss; and The collaborative filtering model is optimized at least by minimizing the cross-system contrastive prediction loss.

14. The apparatus according to claim 13, wherein, The generated candidate content items include: The candidate content item representation is generated based on the knowledge graph corresponding to the candidate content item.

15. The apparatus according to claim 14, wherein, The generation of the candidate content item includes: Identify a set of neighboring nodes adjacent to the candidate content item and a set of edges corresponding to the set of neighboring nodes from the knowledge graph; and The candidate content item representation is generated based on at least one of the following: 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 the importance of the set of neighbor nodes to the candidate content item.

16. The apparatus according to claim 13, wherein, Generating user interest representations includes: The user interest representation is generated by aggregating the set of historical content item representations using the set of overall interest representations.

17. A computer program product for content recommendation based on graph-enhanced collaborative filtering, comprising a computer program executed by at least one processor for: Generate a candidate content item representation for the candidate content item; Generate a set of historical content item representations corresponding to a set of historical content items for the target user; A set of overall interest representations for all users is generated based on a set of meta-interests, where each meta-interest represents an interest element. The user interest representation of the target user is generated based on the set of historical content item representations and the set of overall interest representations; as well as The 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. The computer program product is embodied in the collaborative filtering model, and the training of the collaborative filtering model includes: Using the collaborative filtering model and the content-based filtering model, cross-system contrastive learning is employed to generate a cross-system contrastive prediction loss; and The collaborative filtering model is optimized at least by minimizing the cross-system contrastive prediction loss.

Images