Personalized news recommendation method and system based on interactive heterogeneous graph neural network

Abstract

The application provides a personalized news recommendation method and system based on an interactive heterogeneous graph neural network, relates to the technical field of news recommendation, and comprises the following steps: obtaining user historical news and candidate news; training an interactive heterogeneous graph neural network; inputting the user historical news and the candidate news into the interactive heterogeneous graph neural network to output a prediction score result of user clicks on the candidate news; performing word embedding and entity embedding on the user historical news and the candidate news respectively; obtaining user historical news representation with semantic interaction information through a CNN layer and a Transformer layer; forming a heterogeneous user graph through the heterogeneous gated graph neural network, learning candidate news representation through an enhanced related attention module; and obtaining a prediction score through negative sampling technology from user representation and news representation. The application improves the accuracy of personalized news recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of news recommendation, and particularly relates to an individualized news recommendation method and system based on an interactive heterogeneous graph neural network. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Individualized news recommendation is a process of predicting the relevance between news articles and users and recommending news articles to users to meet their information needs. Compared with other fields of recommendation, news recommendation has greater challenges because the natural characteristics of news articles such as timeliness and novelty result in their highly dynamic characteristics, for example, the relevance of news articles changes rapidly. Such a system not only helps users filter a large amount of news, but also actively recommends news items that meet the real interests of users by using individualized algorithms, thereby greatly meeting the information needs of users.

[0004] In recent years, the technology based on graph neural networks has developed rapidly, and the research enthusiasm about graph neural networks in the field of deep learning is growing, and graph neural networks have become a research hotspot of major deep learning top conferences. The excellent ability of graph neural networks in processing unstructured data enables them to make new breakthroughs in network data analysis, recommendation systems, physical modeling, natural language processing and combinatorial optimization problems on graphs, and bring new opportunities for enhancing content-based solutions in news recommendation, because these technologies have strong ability to process structured data and can capture the hidden interest bias in user reading history. The latest graph neural network news recommendation model has established an effective feature interaction model between news and user graphs, thereby realizing accurate news-user representation matching.

[0005] However, the existing news recommendation method extracts semantic information from users and candidate news respectively, ignoring the semantic interaction information between users and candidate news. In addition, the existing news recommendation method only uses the same node type for message passing, ignoring the different features and topological structures between different node types. The previous news recommendation method learns news representation through text representation, ignoring the semantic association information between entities and text. These places still need to be improved. SUMMARY

[0006] To overcome the above deficiencies of the prior art, the application provides a personalized news recommendation method and system based on an interactive heterogeneous graph neural network, a collaborative fusion module is designed to obtain semantic interaction information between user historical news and candidate news, and the module uses affinity matrix to calculate the correlation weight between different words.

[0007] To achieve the above object, one or more embodiments of the application provide the following technical solutions:

[0008] The first aspect of the application provides a personalized news recommendation method based on an interactive heterogeneous graph neural network.

[0009] A personalized news recommendation method based on an interactive heterogeneous graph neural network comprises:

[0010] Obtaining user historical news and candidate news;

[0011] Training an interactive heterogeneous graph neural network;

[0012] Inputting the user historical news and the candidate news into the interactive heterogeneous graph neural network to output a predicted score result of user clicking the candidate news;

[0013] The structure of the interactive heterogeneous graph neural network comprises a word and entity embedding layer, a collaborative fusion module, a heterogeneous gated graph neural network, an enhanced correlation attention module and a predicted score layer;

[0014] The word embedding and entity embedding of the user historical news and the candidate news are performed through the word and entity embedding layer, and the word embedding vector sequence and the entity embedding vector sequence of the user historical news and the candidate news are obtained respectively;

[0015] The word embedding vector sequence of the user historical news and the candidate news is learned by a CNN layer and a Transformer layer through the collaborative fusion module to obtain user historical news representation with semantic interaction information;

[0016] The user historical news representation is formed into a heterogeneous user graph according to time adjacency through the heterogeneous gated graph neural network, and the heterogeneous graph is encoded by a multi-layer GGNN to obtain user representation;

[0017] The enhanced related attention module is used for learning candidate news representation by using attention mechanism on the word embedding vector sequence and the entity embedding vector sequence of the candidate news.

[0018] The second aspect of the application provides a personalized news recommendation system based on an interactive heterogeneous graph neural network.

[0019] The personalized news recommendation system based on the interactive heterogeneous graph neural network comprises:

[0020] The acquisition module is configured to acquire user historical news and candidate news.

[0021] The training module is configured to train the interactive heterogeneous graph neural network.

[0022] The output module is configured to input the user historical news and the candidate news into the trained interactive heterogeneous graph neural network, and output a prediction score result of user clicks on the candidate news.

[0023] The structure of the interactive heterogeneous graph neural network comprises a word and entity embedding layer, a collaborative fusion module, a heterogeneous gated graph neural network, an enhanced related attention module and a prediction score layer.

[0024] The word and entity embedding layer is used for performing word embedding and entity embedding on the user historical news and the candidate news respectively, and obtaining a word embedding vector sequence and an entity embedding vector sequence of the user historical news and the candidate news respectively.

[0025] The collaborative fusion module is used for learning local and global representations of the word embedding vector sequence of the user historical news and the candidate news through a CNN layer and a Transformer layer, and obtaining user historical news representation with semantic interaction information.

[0026] The heterogeneous gated graph neural network is used for forming a heterogeneous user graph according to the time adjacency of the user historical news representation, and encoding the heterogeneous graph by using a multi-layer GGNN to obtain user representation.

[0027] The enhanced related attention module is used for learning candidate news representation by using attention mechanism on the word embedding vector sequence and the entity embedding vector sequence of the candidate news; and the prediction score layer is used for obtaining a prediction score by using a negative sampling technique on the user representation and the news representation.

[0028] The third aspect of the application provides a computer device, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the processor implements the steps in the method according to the first aspect of the application when executing the program.

[0029] The fourth aspect of the present application provides a computer readable storage medium having stored thereon a program which, when executed by a processor, implements the steps in the method according to the first aspect of the present application.

[0030] The fifth aspect of the present application provides a computer program product comprising instructions which, when executed on a computer, cause the computer program to be executed by a processor to implement the steps in the method according to the first aspect of the present application.

[0031] The above one or more technical solutions have the following beneficial effects:

[0032] An interactive heterogeneous graph neural network including a word and entity embedding layer, a collaborative fusion module, a heterogeneous gate graph neural network, an enhanced related attention module and a prediction score module is designed in the present application; wherein the collaborative fusion module uses an affinity matrix to calculate the correlation weight between different words, fully interacts the user historical news with the candidate news, and finally obtains the user representation by using semantic interaction information, solving the problem of neglecting hidden semantic information in separate modeling.

[0033] The present application constructs a new user heterogeneous graph through the heterogeneous gate graph neural network, uses the gate graph neural network with GRU to map different node types for message passing to obtain high-order information hidden in the user graph, and finally obtains the final user representation by using the high-order information.

[0034] The present application aggregates news text representation and entity representation through the enhanced related attention module to capture the semantic correlation between entities and texts to obtain semantic correlation information, and uses a multi-head self-attention network to unify and aggregate the two representations into news representation to obtain the final news representation with entity relationship information.

[0035] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0036] The accompanying drawings, which form a part of the present application, are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification, illustrate embodiments of the present application and serve to explain the present application, and do not limit the present application in any manner.

[0037] Figure 1 A model flowchart of the first embodiment of the present application;

[0038] Figure 2 A collaborative fusion module schematic diagram of the first embodiment of the present application;

[0039] Figure 3 A fusion related attention module schematic diagram of the first embodiment of the present application. Detailed Implementation

[0040] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0042] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0043] Example 1

[0044] This embodiment discloses a personalized news recommendation method based on an interactive heterogeneous graph neural network, including:

[0045] Retrieve user's historical news and candidate news;

[0046] Training an interactive heterogeneous graph neural network;

[0047] The user's historical news and candidate news are input into an interactive heterogeneous graph neural network, which outputs the predicted score of the candidate news clicked by the user.

[0048] The structure of the interactive heterogeneous graph neural network includes: a word and entity embedding layer, a collaborative fusion module, a heterogeneous gated graph neural network, an enhanced relevance attention module, and a prediction score layer;

[0049] Through the word and entity embedding layer, word embedding and entity embedding are performed on the user's historical news and candidate news respectively, to obtain the word embedding vector sequence and entity embedding vector sequence of the user's historical news and candidate news respectively.

[0050] Through the collaborative fusion module, the word embedding vector sequences of the user's historical news and candidate news are used to learn local and global representations through CNN layers and Transformer layers to obtain a user historical news representation with semantic interaction information.

[0051] The heterogeneous gated graph neural network is used to form a heterogeneous user graph based on the temporal adjacency of the user's historical news representation. The heterogeneous graph is then encoded using a multi-layer GGNN to obtain the user representation.

[0052] The enhanced relevant attention module learns candidate news representations by using an attention mechanism to learn the word embedding vector sequence and entity embedding vector sequence of candidate news; the prediction score layer obtains prediction scores by using negative sampling techniques to obtain user representations and news representations.

[0053] To make the embodiment clearer, the personalized news recommendation process based on the interactive heterogeneous graph neural network can be specifically described as follows:

[0054] As shown in Figure 1 The interactive heterogeneous graph neural network in the embodiment includes a word and entity embedding layer, a collaborative fusion module, a heterogeneous gated graph neural network, an enhanced relevant attention module, and a predictive score layer.

[0055] The input end of the word embedding layer is used to input the user historical news text and the candidate news text, and the input end of the entity embedding layer is used to input the user historical news entity and the candidate news entity.

[0056] The output end of the word embedding layer is connected with the input end of the collaborative fusion module.

[0057] The output end of the entity embedding layer is connected with the input end of the enhanced relevant attention module.

[0058] The output end of the collaborative fusion module is connected with the output end of the heterogeneous gated graph neural network.

[0059] The output ends of the heterogeneous graph neural network and the enhanced relevant attention module are both connected with the input end of the predictive score module.

[0060] The output end of the predictive score module is used to output the score of the user clicking the candidate news.

[0061] In the embodiment, first, the clicked news text data of a known user is obtained as a training set; then the training set is input into the interactive heterogeneous graph neural network to train the network to obtain the trained interactive heterogeneous graph neural network.

[0062] Secondly, the user historical news and the candidate news are input into the trained interactive heterogeneous graph neural network, and the predictive score result of the user clicking the candidate news is output.

[0063] The specific steps are as follows:

[0064] Step 1, the word embedding and entity embedding of the user historical news and the candidate news are performed through the word and entity embedding layer, and the word embedding vector sequence and the entity embedding vector sequence of the user historical news and the candidate news are obtained respectively.

[0065] Step 1-1, converting the user historical news text sequence and the candidate news text sequence into low-dimensional word embedding vector sequences and respectively through a word embedding layer.

[0066] wherein d represents the dimension of the word embedding vector, n represents the number of words, and a pre-trained word embedding model is used to initialize the word embedding layer.

[0067] Step 1-2, learning entity representations of the user historical news entities and the candidate news entities through an entity embedding layer, wherein the entity representations contain entity relationship information in the knowledge graph, and converting the entity representations into entity embedding sequences and .

[0068] wherein E represents the number of entities.

[0069] The entity embedding layer is initialized using a pre-trained entity embedding .

[0070] Step 2, learning local and global representations of the word embedding vector sequences of the user historical news and the candidate news through a CNN layer and a Transformer layer in the collaborative fusion module to obtain user historical news representations with semantic interaction information.

[0071] As shown in Figure 2 , the collaborative fusion module includes a CNN layer and a Transformer layer (CNN Layer and Transformer Layer), a representation sum layer (Representation Sum Layer), and a fusion interaction layer (Fusion Interaction Layer).

[0072] Step 2-1, learning local and global representations of the word vector sequences and of each user historical news through a CNN layer and a Transformer layer, respectively.

[0073] Step 2-1-1, applying and to the CNN layer to generate local vector sequences and , respectively. ​

[0074] For example, in the CNN layer, it is represented as:

[0075] ;

[0076] wherein, and represent the parameters of the CNN filter, respectively representing the kernel and the bias, and T represents the number of words in the text. A nonlinear function GeLU is used as the activation function instead of the Relu function, thereby improving the accuracy.

[0077] Therefore, the output of the CNN layer is a local vector sequence with local context , represented as . Among them, represents the dimension of the user representation, and T represents the number of words in the text.

[0078] Step 2-1-2, apply the vector sequence and to the Transformer network to generate global vector sequences and .

[0079] For example, its calculation formula is:

[0080] ;

[0081] wherein, , represented as .

[0082] The output of the Transformer layer is a global vector sequence with global context , represented as . Among them, represents the dimension of the user representation, and T represents the number of words in the text.

[0083] Local context can reflect the specific behavior and interest points of the user at a specific moment, such as the content and related topics of a certain piece of science and technology news that the user is reading. Global context covers the user's long-term browsing history, search records and other information, representing the overall interest of the user. The combination of the two can more comprehensively understand the user's interest, both capturing the user's temporary interest at the moment and taking into account their long-term stable interest areas, thereby mining potential interests.

[0084] Step 2-2, add each word vector in the local vector sequence and the global vector sequence generated from the user's historical news text and the candidate news text to obtain their unified representation.

[0085] ​Step 2-2-1, add each word vector in the local vector sequence and the global vector sequence of the generated user history news text to obtain their unified representation , denoted as .

[0086] Step 2-2-2, use the same method in step 2-2-1 to obtain the unified representation from the local vector sequence and the global vector sequence of the generated candidate news text , denoted as .

[0087] Step 2-3, use the affinity matrix to obtain the semantic interaction information between the user history news and the candidate news, and learn the representations of the news texts and .

[0088] Step 2-3-1, use the affinity matrix to represent the weights between different words in the news texts and :

[0089] ;

[0090] where is a trainable weight, and T represents the number of words in the text.

[0091] Using the affinity matrix to calculate the weights can guide the model to learn the key information in the data more effectively, reducing the attention to irrelevant or noise information, thereby improving the training speed and convergence performance of the model. For example, in natural language processing tasks, by calculating the weights between words or phrases through the affinity matrix, the model can capture semantic information in the text more quickly, speeding up the training process of the model, and improving the performance of the model in text generation, news recommendation and other tasks.

[0092] Step 2-3-2, calculate the word attention vector of according to :

[0093] ;

[0094] where is the attention query vector, and are trainable matrices. represents the dimension of the attention query vector, represents the dimension of the user representation. represents the ReLU function.

[0095] Step 2-3-3, learning each user historical news representation with semantic interaction information :

[0096] ;

[0097] ;

[0098] where, is the weight value of the i-th word of the text . represents the i-th vector in , represents the i-th word attention vector, represents the j-th word attention vector.

[0099] Step 3, forming a heterogeneous user graph according to the time adjacency of the user historical news representation by the heterogeneous gated graph neural network, and encoding the heterogeneous graph by a multi-layer GGNN to obtain the user representation.

[0100] Step 3-1, by the heterogeneous gated graph neural network, each news representation obtained in step 2 is taken as a news node in the user representation , and they are connected according to the time adjacency to form a user graph, where U represents the number of user historical news.

[0101] In addition, we regard the entity as another type of node, integrate the entity sequence into the user graph, and connect each news node in the user graph with its entity node to finally form a user heterogeneous graph G.

[0102] Step 3-2, selecting multi-hop news neighbors, for each hop, we select neighbors. We map the entity to the text space to obtain :

[0103] ;

[0104] where, is a trainable matrix, is a bias.

[0105] Step 3-3, encoding the heterogeneous graph by a multi-layer GGNN to obtain the user representation :

[0106] ;

[0107] where,t i denotes a heterogeneous graph node.

[0108] Step 3-4, the GGNN network uses a Gated Recurrent Unit (GRU) to capture sequence information. In this embodiment, the news text node and the entity node use similar calculation equations to update the intermediate nodes of each layer of the GGNN, taking the news text node as an example:

[0109] ;

[0110] ;

[0111] ;

[0112] wherein, is initialized using , denotes the neighbor nodes of node i in the user heterogeneous graph, denotes the current layer number of the GGNN, denotes a trainable matrix, denotes the intermediate node updated by the GGNN, denotes the neighbor nodes of the i-th layer.

[0113] Step 4, the enhanced related attention module is used to learn the news representation by using the attention mechanism on the word embedding vector sequence and the entity embedding vector sequence of the candidate news.

[0114] As shown in Figure 3 , the enhanced related attention module includes a Multi-Head Self-Attention Layer and a Representation Pooling Layer.

[0115] Step 4-1, the Multi-Head Self-Attention Layer learns the news representation by using the word embedding representation and the entity embedding representation of the candidate news.

[0116] The word embedding sequence of the candidate news is input into the text Multi-Head Self-Attention (MSA) layer to learn the text representation :

[0117] ;

[0118] ​​​;

[0119] ;

[0120] wherein, is a connection function, is a trainable matrix for output connection. For , the matrix , , maps to query, key and value vectors. Att(-) is used to obtain the vector of each head by using the dot product method to calculate the sum of value vectors, denotes the dimension of the key vector.

[0121] At the same time, the entity embedding sequence of the candidate news is passed through the text multi-head self-attention layer to learn the entity representation .

[0122] Step 4-2, the representation pooling layer, pools the text representation and the entity representation into , and then learns the candidate news representation from by using the attention mechanism:

[0123] ;

[0124] ;

[0125] wherein, is the transpose of the query vector, is a trainable matrix, is the i-th connection vector in , denotes the weight vector of the i-th connection vector of , and j represents the parameter of summation.

[0126] Step 5, by the prediction score layer, the user representation and the news representation are obtained by the negative sampling technique to obtain the prediction score.

[0127] The prediction score layer, in the training data set, adopts the negative sampling technique, records each clicked news as a positive sample , and records the unclicked news as a negative sample . The inner product of each candidate news representation and the user representation in all samples is calculated, and then the click score is calculated by using the softmax function. Wherein, it includes one positive sample and K negative samples:

[0128] ;

[0129] log-likelihood loss of positive samples of the training sample set optimization is performed:

[0130]

[0131] As the log-likelihood loss is continuously optimized, the prediction probability of the model for the positive samples will become closer and closer to the true situation. Through the optimization of the log-likelihood loss, the model can more accurately identify the user's interest in the news, and the probability of correctly classifying the news recommendation as a positive sample will be significantly improved.

[0132] For an unbalanced data set (the number of positive samples and negative samples is greatly different), optimizing the positive samples by the log-likelihood loss helps to prevent the model from paying too much attention to the larger number of negative samples, thereby balancing the classification ability of the model for positive and negative samples to some extent, and improving the recall rate and precision of positive samples.

[0133] In the experiment, the method is applied to two authoritative news data sets MIND and Adressa. In the following experiments, the two data sets are divided into training set, test set and validation set three subsets, and compared with a series of baseline models to verify the effectiveness of the model.

[0134] The model in this embodiment is compared with other baseline models on the data sets MIND and Adressa, and the experimental results have obvious improvement. The semantic interaction information between the user's historical news and the candidate news is obtained by using the affinity matrix in the collaborative fusion module, which solves the problem of ignoring hidden semantic information in separate modeling. In addition, by constructing a new user heterogeneous graph, using a gated graph neural network with GRU for user graph message passing, and using a heterogeneous gated graph neural network to learn a user representation with high-order information. In addition, the enhanced related attention module is used to combine the text representation and the entity representation, and the two representations are unified and aggregated into news representation by using a multi-head self-attention network.

[0135] Embodiment Two

[0136] The purpose of this embodiment is to provide an interactive heterogeneous graph neural network-based personalized news recommendation system, comprising:

[0137] The acquisition module is configured to acquire user historical news and candidate news;

[0138] The training module is configured to train the interactive heterogeneous graph neural network;

[0139] The output module is configured to input the user historical news and the candidate news into the trained interactive heterogeneous graph neural network, and output the prediction score result of the user clicking the candidate news;

[0140] ​The structure of the interactive heterogeneous graph neural network comprises a word and entity embedding layer, a collaborative fusion module, a heterogeneous gated graph neural network, an enhanced related attention module and a prediction score layer.

[0141] The word and entity embedding layer is configured to perform word embedding and entity embedding on the user historical news and the candidate news respectively to obtain word embedding vector sequences and entity embedding vector sequences of the user historical news and the candidate news.

[0142] The collaborative fusion module is configured to learn local and global representations of the word embedding vector sequences of the user historical news and the candidate news through a CNN layer and a Transformer layer to obtain user historical news representations with semantic interaction information.

[0143] The heterogeneous gated graph neural network is configured to form a heterogeneous user graph according to the user historical news representations, encode the heterogeneous graph through a multi-layer GGNN, and obtain user representations.

[0144] The enhanced related attention module is configured to learn candidate news representations from the word embedding vector sequences and the entity embedding vector sequences of the candidate news through an attention mechanism.

[0145] Embodiment Three

[0146] The embodiment aims to provide a computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0147] Embodiment Four

[0148] The embodiment aims to provide a computer readable storage medium.

[0149] A computer readable storage medium having a computer program stored thereon, wherein the program is executable on a processor to perform the steps of the above method.

[0150] Embodiment Five

[0151] The embodiment aims to provide a computer program product comprising instructions which, when executed on a computer, cause the computer to perform the method and functions involved in any of the above embodiments.

[0152] The steps involved in the apparatus of the above embodiments correspond to the method embodiment one, and the specific implementation can refer to the relevant description part of embodiment one. The term "computer readable storage medium" should be understood as including a single medium or multiple media of one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying a set of instructions for execution by a processor and causing the processor to perform any of the methods in the present application.

[0153] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be respectively made into each integrated circuit module, or a plurality of modules or steps among them can be made into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0154] Although the specific embodiments of the present application are described above in combination with the drawings, it is not a limitation on the scope of protection of the present application, and those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A personalized news recommendation method based on interactive heterogeneous graph neural networks, characterized in that, include: Retrieve user's historical news and candidate news; Training an interactive heterogeneous graph neural network; The user's historical news and candidate news are input into an interactive heterogeneous graph neural network, which outputs the predicted score of the candidate news clicked by the user. The structure of the interactive heterogeneous graph neural network includes: a word and entity embedding layer, a collaborative fusion module, a heterogeneous gated graph neural network, an enhanced relevance attention module, and a prediction score layer; Through the word and entity embedding layer, word embedding and entity embedding are performed on the user's historical news and candidate news respectively, to obtain the word embedding vector sequence and entity embedding vector sequence of the user's historical news and candidate news respectively. Through the collaborative fusion module, the word embedding vector sequences of the user's historical news and candidate news are processed through CNN and Transformer layers to learn local and global representations, thereby obtaining a user historical news representation with semantic interaction information; specifically: The word vector sequences of each user's historical news and the word vector sequences of candidate news are used to learn local and global representations through CNN layers and Transformer layers, respectively. The word vectors in the local and global vector sequences of the generated user history news text and candidate news text are added together to obtain their unified representation; An affinity matrix is ​​used to obtain semantic interaction information between a user's historical news and candidate news. The heterogeneous gated graph neural network (GGNN) is used to form a heterogeneous user graph based on the temporal adjacency of user historical news representations. A multi-layer GGNN is then used to encode the heterogeneous user graph to obtain the user representation; specifically: By using a heterogeneous gated graph neural network, each news representation is treated as a news node in the user representation, and they are connected according to temporal adjacency to form a user graph. Treating entities as another type of node, integrating entity sequences into the user graph, and connecting each news node in the user graph with its entity nodes, ultimately forming a heterogeneous user graph; Selecting multi-hop news neighbors maps entities to text space; A multi-layer GGNN is used to encode the heterogeneous user graph to obtain user representations; The enhanced relevant attention module learns news representations by using an attention mechanism to learn word embedding vector sequences and entity embedding vector sequences of candidate news; the prediction score layer obtains prediction scores by using negative sampling techniques to obtain user representations and news representations.

2. The personalized news recommendation method based on an interactive heterogeneous graph neural network as described in claim 1, characterized in that, The process of performing word embedding and entity embedding on the user's historical news and candidate news respectively, to obtain word embedding vector sequences and entity embedding vector sequences for the user's historical news and candidate news, is as follows: The user's historical news text sequence and candidate news text sequence are transformed into low-dimensional word embedding vector sequences through a word embedding layer; Then, the entity representations of the user's historical news entities and candidate news entities are learned through the entity embedding layer and converted into entity embedding sequences. Among them, those using pre-trained The model initializes the word embedding layer, and the entity embedding layer uses pre-trained... Entity embedding To initialize.

3. The personalized news recommendation method based on an interactive heterogeneous graph neural network as described in claim 1, characterized in that, The process of learning news representations using attention mechanisms by combining word embedding vector sequences and entity embedding vector sequences of candidate news items specifically involves: The word embedding sequence of candidate news is passed through a text multi-head self-attention layer to learn text representation; The entity embedding sequence of candidate news is passed through a text multi-head self-attention layer to learn entity representations; Then, the text representation and entity representation are pooled, and then an attention mechanism is used to learn candidate news representations from the pooled representations.

4. The personalized news recommendation method based on an interactive heterogeneous graph neural network as described in claim 1, characterized in that, Through the aforementioned prediction score layer, user representations and news representations are used to obtain prediction scores via negative sampling techniques, specifically as follows: The negative sampling technique is used to record each clicked news item as a positive sample and the unclicked news item as a negative sample. The inner product of the candidate news representation and the user representation in all samples is calculated, and then the click score is calculated using the softmax function.

5. A personalized news recommendation system based on an interactive heterogeneous graph neural network, characterized in that, include: The acquisition module is configured to acquire users' historical news and candidate news. The training module is configured to train an interactive heterogeneous graph neural network. The output module is configured to input the user's historical news and candidate news into a trained interactive heterogeneous graph neural network and output the predicted score of the candidate news clicked by the user. The structure of the interactive heterogeneous graph neural network includes: a word and entity embedding layer, a collaborative fusion module, a heterogeneous gated graph neural network, an enhanced relevance attention module, and a prediction score layer; Through the word and entity embedding layer, word embedding and entity embedding are performed on the user's historical news and candidate news respectively, to obtain the word embedding vector sequence and entity embedding vector sequence of the user's historical news and candidate news respectively. Through the collaborative fusion module, the word embedding vector sequences of the user's historical news and candidate news are processed through CNN and Transformer layers to learn local and global representations, thereby obtaining a user historical news representation with semantic interaction information; specifically: The word vector sequences of each user's historical news and the word vector sequences of candidate news are used to learn local and global representations through CNN layers and Transformer layers, respectively. The word vectors in the local and global vector sequences of the generated user history news text and candidate news text are added together to obtain their unified representation; An affinity matrix is ​​used to obtain semantic interaction information between a user's historical news and candidate news. The heterogeneous gated graph neural network (GGNN) is used to form a heterogeneous user graph based on the temporal adjacency of user historical news representations. A multi-layer GGNN is then used to encode the heterogeneous user graph to obtain the user representation; specifically: By using a heterogeneous gated graph neural network, each news representation is treated as a news node in the user representation, and they are connected according to temporal adjacency to form a user graph. Treating entities as another type of node, integrating entity sequences into the user graph, and connecting each news node in the user graph with its entity nodes, ultimately forming a heterogeneous user graph; Selecting multi-hop news neighbors maps entities to text space; A multi-layer GGNN is used to encode the heterogeneous user graph to obtain user representations; The enhanced relevant attention module learns news representations by using an attention mechanism to learn word embedding vector sequences and entity embedding vector sequences of candidate news; the prediction score layer obtains prediction scores by using negative sampling techniques to obtain user representations and news representations.

6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the method described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.

Images