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

Abstract

The invention provides a personalized news recommendation method and system based on an interactive heterogeneous graph neural network, and relates to the technical field of news recommendation, and the method comprises the steps: obtaining historical news and candidate news of a user; training an interactive heterogeneous graph neural network; inputting the historical news and the candidate news of the user into the interactive heterogeneous graph neural network, and outputting a predicted score result of the candidate news clicked by the user; performing word embedding and entity embedding on the historical news and the candidate news of the user respectively; obtaining user historical news representation with semantic interaction information through a CNN layer and a Transform layer; forming a heterogeneous user graph through the heterogeneous gated graph neural network, and learning candidate news representation by enhancing a related attention module; and obtaining prediction scores of the user characterization and the news characterization through a negative sampling technology. According to the invention, the accuracy of personalized news recommendation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of news recommendation, and in particular to a personalized news recommendation method and system based on an interactive heterogeneous graph neural network. Background Art

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

[0003] Personalized news recommendation is the process of recommending news articles to users to meet their information needs by predicting the relevance between them. Compared with recommendations in other fields, news recommendation is more challenging because news articles are highly dynamic due to their natural characteristics such as timeliness and novelty. For example, the relevance of news articles can change rapidly. Such a system can not only help users filter massive amounts of news, but also use personalized algorithms to proactively recommend news items that meet users' real interests, thereby greatly satisfying users' information needs.

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

[0005] However, existing news recommendation methods extract semantic information from users and candidate news separately, ignoring the semantic interaction information between users and candidate news. In addition, existing news recommendation methods only use the same node type for message passing, ignoring the different features and topological structures between different node types. Previous news recommendation methods learn news representation through text representation, ignoring the entity relationship and semantic association information between texts. These areas still need to be improved. Summary of the invention

[0006] In order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a personalized news recommendation method and system based on an interactive heterogeneous graph neural network, and designs a collaborative fusion module to obtain the semantic interaction information between the user's historical news and the candidate news. The module uses an affinity matrix to calculate the correlation weights between different words. At the same time, a heterogeneous gated graph neural network is designed to map different node types for message passing, so that high-order semantic information of the graph can be obtained. In addition, an enhanced correlation attention module is designed to aggregate text representation and entity representation into a unified representation, thereby capturing the semantic correlation between entity relationships and text.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: A first aspect of the present invention provides a personalized news recommendation method based on an interactive heterogeneous graph neural network.

[0008] A personalized news recommendation method based on interactive heterogeneous graph neural network, comprising: Get user historical news and candidate news; Training interactive heterogeneous graph neural networks; Input the user's historical news and candidate news into the interactive heterogeneous graph neural network, and output the predicted score result of the user clicking on the candidate news; 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 correlation attention module, and a prediction score layer. Through the word and entity embedding layer, the user's historical news and candidate news are respectively embedded in words and embedded in entities, so as to obtain word embedding vector sequences and entity embedding vector sequences 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 passed through the CNN layer and the Transformer layer to learn local and global representations, and obtain the user's historical news representation with semantic interaction information; Through the heterogeneous gated graph neural network, the user's historical news representation is formed into a heterogeneous user graph according to time proximity, and the heterogeneous graph is encoded using a multi-layer GGNN to obtain the user representation; Through the enhanced relevant attention module, the word embedding vector sequence and entity embedding vector sequence of the candidate news are used to learn the candidate news representation using the attention mechanism; through the prediction score layer, the user representation and the news representation are used to obtain the prediction score through the negative sampling technology.

[0009] A second aspect of the present invention provides a personalized news recommendation system based on an interactive heterogeneous graph neural network.

[0010] A personalized news recommendation system based on interactive heterogeneous graph neural network, including: An acquisition module, which is configured to: acquire user historical news and candidate news; A training module configured to train an interactive heterogeneous graph neural network; An output module, which is configured to: input the user's historical news and candidate news into the trained interactive heterogeneous graph neural network, and output the predicted score result of the user clicking on the candidate news; 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 correlation attention module, and a prediction score layer. Through the word and entity embedding layer, the user's historical news and candidate news are respectively embedded in words and embedded in entities, so as to obtain word embedding vector sequences and entity embedding vector sequences 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 passed through the CNN layer and the Transformer layer to learn local and global representations, and obtain the user's historical news representation with semantic interaction information; Through the heterogeneous gated graph neural network, the user's historical news representation is formed into a heterogeneous user graph according to time proximity, and the heterogeneous graph is encoded using a multi-layer GGNN to obtain the user representation; Through the enhanced relevant attention module, the word embedding vector sequence and entity embedding vector sequence of the candidate news are used to learn the candidate news representation using the attention mechanism; through the prediction score layer, the user representation and the news representation are used to obtain the prediction score through the negative sampling technology.

[0011] The third aspect of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in a method as described in the first aspect of the present invention are implemented.

[0012] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in a method as described in the first aspect of the present invention.

[0013] A fifth aspect of the present invention provides a computer program product comprising instructions, which, when running on a computer, enables the computer program to implement the steps in a method as described in the first aspect of the present invention when executed by a processor.

[0014] One or more of the above technical solutions have the following beneficial effects: The present invention designs an interactive heterogeneous graph neural network including a word and entity embedding layer, a collaborative fusion module, a heterogeneous gated graph neural network, an enhanced relevant attention module and a prediction score module; wherein the collaborative fusion module uses an affinity matrix to calculate the relevant weights between different words, fully interacts the user's historical news with the candidate news, and finally uses the semantic interaction information to obtain the user representation, thus solving the problem of ignoring hidden semantic information in separate modeling.

[0015] The present invention constructs a new user heterogeneous graph through a heterogeneous gated graph neural network, uses a gated 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 uses this high-order information to obtain the final user representation.

[0016] The present invention obtains semantic association information by enhancing the relevant attention module and aggregating news text representation and entity representation to capture the semantic correlation between entity relations and texts. The two representations are unified and aggregated into a news representation by using a multi-head self-attention network to obtain the final news representation with entity relationship information.

[0017] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0019] Figure 1 This is a model flow chart of the first embodiment of the present invention; Figure 2 This is a schematic diagram of a collaborative fusion module according to the first embodiment of the present invention; Figure 3 Schematic diagram of the fusion-related attention module of embodiment 1 of the present invention. DETAILED DESCRIPTION

[0020] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0021] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0022] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0023] Embodiment 1 This embodiment discloses a personalized news recommendation method based on an interactive heterogeneous graph neural network, including: Get user historical news and candidate news; Training interactive heterogeneous graph neural networks; Input the user's historical news and candidate news into the interactive heterogeneous graph neural network, and output the predicted score result of the user clicking on the candidate news; 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 correlation attention module, and a prediction score layer. Through the word and entity embedding layer, the user's historical news and candidate news are respectively embedded in words and embedded in entities, so as to obtain word embedding vector sequences and entity embedding vector sequences 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 passed through the CNN layer and the Transformer layer to learn local and global representations, and obtain the user's historical news representation with semantic interaction information; Through the heterogeneous gated graph neural network, the user's historical news representation is formed into a heterogeneous user graph according to time proximity, and the heterogeneous graph is encoded using a multi-layer GGNN to obtain the user representation; Through the enhanced relevant attention module, the word embedding vector sequence and entity embedding vector sequence of the candidate news are used to learn the candidate news representation using the attention mechanism; through the prediction score layer, the user representation and the news representation are used to obtain the prediction score through the negative sampling technology.

[0024] In order to more clearly illustrate this embodiment, the personalized news recommendation process based on the interactive heterogeneous graph neural network can be specifically described as follows: like Figure 1 As shown, the interactive heterogeneous graph neural network in this embodiment includes a word and entity embedding layer (Word and Entity Embedding Layer), a collaborative fusion module (Collaborative Fusion Module), a heterogeneous gated graph neural network (Heterogeneous Gated Graph Neural Network), an Enhanced Relevant Attention Module (EnhancedRelevant Attention Module) and a predictive score layer (Predictive Score Layer).

[0025] The input end of the word embedding layer is used to input the user's historical news text and candidate news text, and the input end of the entity embedding layer is used to input the user's historical news entities and candidate news entities; The output of the word embedding layer is connected to the input of the collaborative fusion module; The output of the entity embedding layer is connected to the input of the enhanced correlation attention module; The output of the collaborative fusion module is connected to the output of the heterogeneous gated graph neural network; The outputs of the heterogeneous graph neural network and the enhanced correlation attention module are connected to the input of the prediction score module; The output of the prediction score module is used to output the score of the user clicking on the candidate news.

[0026] In this embodiment, first, news text data clicked by known users is obtained as a training set; then the training set is input into the interactive heterogeneous graph neural network, and the network is trained to obtain a trained interactive heterogeneous graph neural network.

[0027] Secondly, the user's historical news and candidate news are obtained, and the user's historical news and candidate news are input into the trained interactive heterogeneous graph neural network, and the predicted score results of the user's click on the candidate news are output.

[0028] The specific steps are as follows: Step 1: Through the word and entity embedding layer, the user's historical news and candidate news are respectively embedded in words and embedded in entities to obtain word embedding vector sequences and entity embedding vector sequences of the user's historical news and candidate news respectively.

[0029] Step 1-1: Sequence the user's historical news text and candidate news text sequences After the word embedding layer, they are converted into low-dimensional word embedding vector sequences and .

[0030] in, represents the dimension of the word embedding vector, Represents the number of words, using pre-trained The model initializes the word embedding layer.

[0031] Step 1-2: Learn the user's historical news entities through the entity embedding layer and candidate news entities The entity representation, which contains the entity relationship information in the knowledge graph, is converted into an entity embedding sequence and .

[0032] Where E represents the number of entities.

[0033] The entity embedding layer uses pre-trained Entity Embedding to initialize.

[0034] Step 2: Through the collaborative fusion module, the word embedding vector sequences of the user historical news and candidate news are passed through the CNN layer and the Transformer layer to learn local and global representations to obtain the user historical news representation with semantic interaction information.

[0035] like Figure 2 As shown, 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).

[0036] Step 2-1: Sequence the word vectors of each user's historical news and Local and global representations are learned through CNN layers and Transformer layers respectively.

[0037] Step 2-1-1, respectively and Applied to CNN layers to generate local vector sequences and .

[0038] by For example, it is represented in the CNN layer as: ; in, and Represents the parameters of the CNN filter, representing the kernel and bias respectively, and T represents the number of words in the text. The nonlinear function GeLU is used as the activation function instead of the ReLU function to improve the accuracy.

[0039] Therefore, the output of the CNN layer is a sequence of local vectors with local context , expressed as .in represents the dimension of user representation, and T represents the number of words in the text.

[0040] Step 2-1-2: vector sequence and Applying the Transformer network to generate a global vector sequence and .

[0041] by For example, the calculation formula is: ; in, , expressed as .

[0042] The output of the Transformer layer is a sequence of global vectors with global context , expressed as .in represents the dimension of user representation, and T represents the number of words in the text.

[0043] Local context can reflect the user's specific behavior and interests at a specific moment, such as the content of a piece of technology news that the user is reading and related topics. Global context covers a user's long-term browsing history, search history and other information, representing the user's overall interest tendencies. The combination of the two can provide a more comprehensive insight into user interests, capturing the user's current temporary interests while taking into account their long-term and stable areas of interest, thereby tapping into potential interests.

[0044] Step 2-2: Add the word vectors in the local vector sequence and the global vector sequence of the generated user historical news text and candidate news text to obtain their unified representation.

[0045] Step 2-2-1: Generate the user's historical news text Add the word vectors in the local vector sequence and the global vector sequence to get their unified representation , denoted as .

[0046] Step 2-2-2: Use the same method as in step 2-2-1 to generate candidate news texts The local vector sequence and the global vector sequence of , denoted as .

[0047] Step 2-3: Use affinity matrix to obtain semantic interaction information between user historical news and candidate news, and interactively learn news text and Representation.

[0048] Step 2-3-1: Use affinity matrix Indicates news text and The weights between different words in: ; in, is a trainable weight and T represents the number of words in the text.

[0049] Using affinity matrices to calculate weights can guide the model to learn key information in the data more effectively, reduce attention to irrelevant or noisy information, and thus improve the model's training speed and convergence performance. For example, in natural language processing tasks, using affinity matrices to calculate the weights between words or phrases can allow the model to capture semantic information in the text more quickly, speed up the model's training process, and improve the model's performance in tasks such as text generation and news recommendation.

[0050] Step 2-3-2, according to calculate Word attention vector : ; in, is the attention query vector, and is a trainable matrix. represents the dimension of the attention query vector, Represents the dimension of the user representation. Represents the ReLU function.

[0051] Step 2-3-3: Learn each user’s historical news representation with semantic interaction information : ; ; in, is text The weight value of the i-th word. express The i-th vector in represents the i-th word attention vector, represents the j-th word attention vector.

[0052] Step 3: Through the heterogeneous gated graph neural network, the user's historical news representation is formed into a heterogeneous user graph according to time proximity, and a multi-layer GGNN is used to encode the heterogeneous graph to obtain the user representation.

[0053] Step 3-1: Represent each news obtained in step 2 through a heterogeneous gated graph neural network As user representation and connect them according to time proximity to form a user graph, where U represents the number of historical news for the user.

[0054] In addition, we regard entities as another type of nodes and treat entity sequences as Integrate into the user graph and map each news node in the user graph With its entity node Connect them together to eventually form a user heterogeneous graph G.

[0055] Step 3-2, select multi-hop news neighbors. For each hop, we select neighbors. We will Mapping to text space to obtain : ; in, is a trainable matrix, It's a deviation.

[0056] Step 3-3: Use multi-layer GGNN to encode the heterogeneous graph and obtain user representation : ; in, t i Represents a heterogeneous graph node.

[0057] Step 3-4: The GGNN network uses a gated recurrent unit (GRU) to capture sequence information. In this example, the news text node and entity nodes Use similar calculation equations to update the intermediate nodes of each layer of GGNN to update the news text nodes For example: ; ; ; in, use initialization, represents the neighbor nodes of node i in the user heterogeneous graph, Indicates the current number of layers of GGNN, represents the trainable matrix, Indicates that GGNN The intermediate nodes of the layer update, Indicates Neighbor nodes of the layer.

[0058] Step 4: Through the enhanced relevant attention module, the word embedding vector sequence and entity embedding vector sequence of the candidate news are used to learn news representation using the attention mechanism.

[0059] like Figure 3As shown in the figure, the enhanced related attention module includes a multi-head self-attention layer (Multi-Head Self-Attention Layer) and a representation pooling layer (Representation Pooling Layer).

[0060] Step 4-1: The multi-head self-attention layer is represented by word embedding of candidate news and entity embedding representation Let’s learn news representation.

[0061] Candidate News The word embedding sequence Learn text representation through the text multi-head self-attention (MSA) layer : ; ; ; in, is the connection function, is a trainable matrix of output connections. For example, the matrix , , Will Mapped to query, key, and value vectors. Att(-) is used to obtain the vector for each head by calculating the sum of the value vectors using the dot product method. Indicates the dimension of the key vector. 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 .

[0062] Step 4-2: The representation pooling layer pools the text representation and entity representation into , and then use the attention mechanism from Learning candidate news representations : ; ; in, is the transpose of the query vector, is a trainable matrix, yes The i-th connection vector in , express is the weight vector of the i-th connection vector, and j represents the parameter of the summation.

[0063] Step 5: Through the prediction score layer, the user representation and the news representation are subjected to negative sampling technology to obtain a prediction score.

[0064] The prediction score layer uses negative sampling technology in the training data set to record each clicked news as a positive sample. , record the unclicked news as negative samples . Represent each candidate news in all samples User Characterization Do the inner product, and then use the softmax function to calculate the click score . Among them, there is one positive sample and K negative samples: ; Positive sample log-likelihood loss for the training sample set To optimize: ; As the log-likelihood loss is continuously optimized, the model's predicted probability for positive samples will get closer and closer to the actual situation. Through log-likelihood loss optimization, the model can more accurately identify the user's interest in news, and the probability of correctly classifying news recommendations as positive samples will be significantly improved.

[0065] For unbalanced data sets (where the number of positive and negative samples is quite different), optimizing the positive samples with log-likelihood loss helps prevent the model from paying too much attention to the large number of negative samples, thereby balancing the model's ability to classify positive and negative samples to a certain extent and improving the recall and precision of positive samples.

[0066] In the experiment, this method is applied to two authoritative news datasets, MIND and Adressa. In the following experiments, these two datasets are divided into three subsets: training set, test set and validation set, and compared with a series of baseline models to verify the effectiveness of the model.

[0067] The model in this embodiment is compared with other baseline models on the datasets MIND and Adressa, and the experimental results show significant improvements. The semantic interaction information between the user's historical news and the candidate news is obtained through the affinity matrix using 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, a gated graph neural network with GRU is used for user graph message passing, and a heterogeneous gated graph neural network is used to learn user representations with high-order information. In addition, the text representation is combined with the entity representation using the enhanced relevant attention module, and the two representations are unified and aggregated into a news representation by using a multi-head self-attention network.

[0068] Embodiment 2 The purpose of this embodiment is to provide a personalized news recommendation system based on an interactive heterogeneous graph neural network, including: An acquisition module, which is configured to: acquire user historical news and candidate news; A training module configured to train an interactive heterogeneous graph neural network; An output module, which is configured to: input the user's historical news and candidate news into the trained interactive heterogeneous graph neural network, and output the predicted score result of the user clicking on the candidate news; 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 correlation attention module, and a prediction score layer. Through the word and entity embedding layer, the user's historical news and candidate news are respectively embedded in words and embedded in entities, so as to obtain word embedding vector sequences and entity embedding vector sequences 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 passed through the CNN layer and the Transformer layer to learn local and global representations, and obtain the user's historical news representation with semantic interaction information; Through the heterogeneous gated graph neural network, the user's historical news representation is formed into a heterogeneous user graph according to time proximity, and the heterogeneous graph is encoded using a multi-layer GGNN to obtain the user representation; Through the enhanced relevant attention module, the word embedding vector sequence and entity embedding vector sequence of the candidate news are used to learn the candidate news representation using the attention mechanism; through the prediction score layer, the user representation and the news representation are used to obtain the prediction score through the negative sampling technology.

[0069] Embodiment 3 The purpose of this embodiment is to provide a computer device, including 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.

[0070] Embodiment 4 The purpose of this embodiment is to provide a computer-readable storage medium.

[0071] A computer-readable storage medium stores a computer program, which executes the steps of the above method when executed by a processor.

[0072] Embodiment 5 The purpose of this embodiment is to provide a computer program product containing instructions, which, when running on a computer, enables the computer to execute the methods and functions involved in any of the above embodiments. The steps involved in the apparatus of the above embodiment correspond to the method embodiment 1, and the specific implementation method can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0073] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0074] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A personalized news recommendation method based on interactive heterogeneous graph neural network, characterized in that: include: Get user historical news and candidate news; Training interactive heterogeneous graph neural networks; Input the user's historical news and candidate news into the interactive heterogeneous graph neural network, and output the predicted score result of the user clicking on the candidate news; 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 correlation attention module, and a prediction score layer. Through the word and entity embedding layer, the user's historical news and candidate news are respectively embedded in words and embedded in entities, so as to obtain word embedding vector sequences and entity embedding vector sequences 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 passed through the CNN layer and the Transformer layer to learn local and global representations, and obtain the user's historical news representation with semantic interaction information; Through the heterogeneous gated graph neural network, the user's historical news representation is formed into a heterogeneous user graph according to time proximity, and the heterogeneous graph is encoded using a multi-layer GGNN to obtain the user representation; Through the enhanced relevant attention module, the word embedding vector sequence and entity embedding vector sequence of the candidate news are used to learn news representation using the attention mechanism; through the prediction score layer, the user representation and the news representation are used to obtain the prediction score through the negative sampling technology.

2. A personalized news recommendation method based on an interactive heterogeneous graph neural network as claimed in claim 1, characterized in that: The word embedding and entity embedding are performed on the user's historical news and the candidate news respectively, and the word embedding vector sequence and entity embedding vector sequence of the user's historical news and the candidate news are obtained respectively, specifically: The user's historical news text sequence and the candidate news text sequence are converted into low-dimensional word embedding vector sequences through the 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, using pre-trained The model initializes the word embedding layer and the entity embedding layer using pre-trained Entity Embedding to initialize.

3. The personalized news recommendation method based on interactive heterogeneous graph neural network according to claim 1, characterized in that: Through the collaborative fusion module, the word embedding vector sequences of the user's historical news and candidate news are passed through the CNN layer and the Transformer layer to learn local and global representations, and the user's historical news representation with semantic interaction information is obtained, specifically: The word vector sequence of each user's historical news and the word vector sequence of candidate news are passed through the CNN layer and the Transformer layer to learn local and global representations respectively; Add the word vectors in the local vector sequence and the global vector sequence of the generated user historical news text and candidate news text to obtain their unified representation; The affinity matrix is ​​used to obtain the semantic interaction information between the user's historical news and the candidate news.

4. The personalized news recommendation method based on interactive heterogeneous graph neural network according to claim 1, characterized in that: The user historical news representation is formed into a heterogeneous user graph according to time proximity, and a multi-layer GGNN is used to encode the heterogeneous graph to obtain the user representation, which is specifically: Through 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 their temporal proximity to form a user graph; Treat entities as another type of nodes, integrate entity sequences into the user graph, and connect each news node in the user graph with its entity node to eventually form a user heterogeneous graph; Select multi-hop news neighbors to map entities into text space; A multi-layer GGNN is used to encode the heterogeneous graph and obtain user representation.

5. The personalized news recommendation method based on interactive heterogeneous graph neural network according to claim 1, characterized in that: The word embedding vector sequence and entity embedding vector sequence of the candidate news are used to learn the news representation by using the attention mechanism, specifically: The word embedding sequence of the candidate news is passed through the text multi-head self-attention layer to learn the text representation; The entity embedding sequence of candidate news is passed through a text multi-head self-attention layer to learn entity representation; The text representation and entity representation are then pooled, and the attention mechanism is used to learn candidate news representations from the pooled representations.

6. The personalized news recommendation method based on interactive heterogeneous graph neural network according to claim 1, characterized in that: Through the prediction score layer, the user representation and news representation are used to obtain prediction scores through negative sampling technology, specifically: Using negative sampling technology, each clicked news is recorded as a positive sample, and each unclicked news is recorded as a negative sample; the inner product of each candidate news representation and the user representation in all samples is performed, and then the softmax function is used to calculate the click score.

7. A personalized news recommendation system based on interactive heterogeneous graph neural network, characterized in that: include: An acquisition module, which is configured to: acquire user historical news and candidate news; A training module configured to train an interactive heterogeneous graph neural network; An output module, which is configured to: input the user's historical news and candidate news into the trained interactive heterogeneous graph neural network, and output the predicted score result of the user clicking on the candidate news; 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 correlation attention module, and a prediction score layer. Through the word and entity embedding layer, the user's historical news and candidate news are respectively embedded in words and embedded in entities, so as to obtain word embedding vector sequences and entity embedding vector sequences 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 passed through the CNN layer and the Transformer layer to learn local and global representations, and obtain the user's historical news representation with semantic interaction information; Through the heterogeneous gated graph neural network, the user's historical news representation is formed into a heterogeneous user graph according to time proximity, and the heterogeneous graph is encoded using a multi-layer GGNN to obtain the user representation; Through the enhanced relevant attention module, the word embedding vector sequence and entity embedding vector sequence of the candidate news are used to learn the candidate news representation using the attention mechanism; through the prediction score layer, the user representation and the news representation are used to obtain the prediction score through the negative sampling technology.

8. 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, the steps of the method described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 6 are performed.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images