Neural network platform novel recommendation method based on node feature enhancement

Abstract

The invention discloses a neural network platform novel recommendation method based on node feature enhancement, and belongs to the field of deep learning, a heterogeneous graph neural network model based on node feature enhancement is constructed, and the model comprises a small-degree node connectivity enhancement module, a path purification module and an inter-path aggregation module; the method specifically comprises the following steps: collecting all novel historical data of all users from a current novel platform, processing the data into a heterogeneous graph form, and carrying out feature conversion; based on a small-degree node connectivity enhancement module, expanding a connection relationship of small-degree nodes; purifying the element paths based on a path purification module; generating a final embedded matrix representation of the node based on an inter-path aggregation module; constructing a loss function optimization training model; and obtaining novel data of the current user, inputting the trained model, and generating a personalized novel recommendation list according to the preference of each user. According to the invention, the accuracy of novel recommendation is remarkably improved, and the novel which is most matched with the preference is recommended to the user.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning, and particularly relates to a novel recommendation method for a neural network platform based on node feature enhancement. Background Art

[0002] Currently, when performing novel recommendations based on traditional heterogeneous graph neural networks, the following problems exist.

[0003] First, the performance problem of low-degree nodes; heterogeneous graph neural networks contain various types of nodes and edges. In the novel recommendation scenario, nodes usually include users, novels, authors, novel platforms, etc., and the edges represent the interaction relationships between users and novels (such as reading, liking, collecting) and the associations between novels and categories, authors. Some nodes may not be able to effectively aggregate sufficient feature information from neighbor nodes due to their fewer connections with other nodes, that is, their degree values are small. For example, some user nodes may have only interacted with a small number of novel nodes, which results in these user nodes being unable to fully obtain the feature information of novels; similarly, some unpopular novel nodes cannot obtain sufficient user preference information due to fewer interactions with users. In contrast, user nodes or popular novel nodes with larger degree values can obtain more neighbor feature information through rich interaction relationships, thus generating more expressive representations. This lack of information aggregation limits the representation ability of low-degree nodes, making them perform poorly in subsequent novel recommendation tasks. This not only affects the performance of individual nodes but may also indirectly affect the feature learning process of the entire graph due to the limitations of information transmission, ultimately reducing the performance of the overall recommendation model.

[0004] Second, in the novel recommendation scenario, improper selection of meta-paths may lead to low model training efficiency and even cause the model to fall into a local optimal solution, unable to achieve the best performance. For example, in a heterogeneous graph neural network mainly based on user-novel-category, some meta-paths may not be able to effectively capture the true preferences of users or may wrongly strengthen the relationships between irrelevant nodes. For example, on a certain novel dataset, when using the combination of meta-paths "user-novel-author-novel-user" and "user-novel-category-novel-user", the accuracy of novel recommendation by the model is lower than that when only using the path "user-novel-author-novel-user". This indicates that it is crucial to reasonably select and optimize meta-paths to avoid introducing redundant or interfering information. Summary of the Invention

[0005] To solve the above problems, the present invention proposes a novel recommendation method for a neural network platform based on enhanced node features, and constructs a heterogeneous graph neural network model HGNN-CERE based on enhanced node features. To improve the feature aggregation ability of the Xiaodu node, HGNN-CERE first uses a node importance algorithm to identify important node information in the graph, and then connects the Xiaodu node to these important nodes to improve the connectivity of the Xiaodu node, so as to achieve the purpose of enhancing the feature aggregation ability of the Xiaodu node. At the same time, in order to avoid the influence of unreliable meta-paths during meta-path aggregation, HGNN-CERE proposes a purification mechanism to offset the influence of unreliable meta-paths. Finally, HGNN-CERE adopts a path-level aggregation strategy to fuse the semantic representations under different paths to generate the final feature embedding.

[0006] The technical solution of the present invention is as follows: A novel recommendation method for a neural network platform based on enhanced node features constructs a heterogeneous graph neural network model HGNN-CERE based on enhanced node features. The model includes a Xiaodu node connectivity enhancement module, a path purification module, and an inter-path aggregation module. The method specifically includes the following steps: Step 1: Collect all the historical novel data of all users from the current novel platform, process it into a heterogeneous graph form, and perform feature transformation; Step 2: The Xiaodu node connectivity enhancement module uses a node importance discovery algorithm to identify key node information in the heterogeneous graph, and then connects the Xiaodu node to the key node information to expand the connection relationship of the Xiaodu node; Step 3: Dynamically adjust the weight of each meta-path based on the path purification module; Step 4: The inter-path aggregation module adopts a path-level aggregation strategy to fuse the feature embeddings of different meta-paths to generate the final embedding matrix representation of the node; Step 5: Construct a loss function to optimize the training model HGNN-CERE; Step 6: Obtain the novel data of the current user, input it into the trained model HGNN-CERE, and generate a personalized novel recommendation list according to the preferences of each user.

[0007] Furthermore, the specific process of the above Step 1 is: Step 1.1. The nodes in the heterogeneous graph represent different entities. Since the historical data contains multiple entities such as novels, authors, and users, the nodes in the heterogeneous graph include multiple types such as novels, authors, users, and novel platforms. The edges in the heterogeneous graph represent the relationships between entities, specifically including reading relationships, creation relationships, and subordination relationships. The heterogeneous graph is a graph composed of multiple types of nodes. According to the historical data, the node feature matrix of each entity is constructed, and all the node feature matrices form the required heterogeneous graph. Among them, the node feature matrix includes a novel node feature matrix, an author node feature matrix, and a user node feature matrix. The novel node feature matrix is composed of all novel information, the author node feature matrix is composed of all author information, and the user node feature matrix is composed of user reading records. Step 1.2. The original node features in the node feature matrix are in different feature spaces, and feature transformation is performed on the nodes: (1); Among them, is the original feature of the th node , is the original feature dimension of the node; is the transformation matrix of a certain type of node, is the node feature dimension after transformation; is the feature vector of the th node after feature transformation.

[0008] Furthermore, the specific process of Step 2 is as follows: Step 2.1. Calculate the degree value of each node in the heterogeneous graph: (2); Among them, represents the adjacency matrix of the heterogeneous graph; is the value at the th row and the th column of the adjacency matrix , indicating whether the th node and the th node are connected. The th row corresponds to the th node, and the th column corresponds to the th node; represents the total number of all nodes; is the degree value of the node ; After obtaining the degree values of all nodes, for each node, traverse its region and use a ranking function to select the nodes with low degrees: (3); Among them, is the list of degree values of all nodes; is the ranking function used to rank the degree values and, after passing through the selection function selects the top nodes as the nodes with low degrees; is the list of nodes with low degrees after selection; Step 2.2: Use betweenness centrality as a measure to judge the importance of nodes: (4); Among them, is the betweenness centrality of node ; is the number of paths passing through node from the th node to the th node ; is the number of shortest paths from the th node to the th node ; Step 2.3: Standardize the betweenness centrality; the heterogeneous graph contains a directed graph and an undirected graph. The relationship between users and novels is directed, forming a directed graph, while the relationship between users and users is undirected, forming an undirected graph; for the undirected graph, the standardized betweenness centrality is: (5); Among them, is the standardized betweenness centrality of node in the undirected graph; is the total number of nodes in the undirected graph; For the directed graph, the standardized betweenness centrality is: (6); Among them, is the standardized betweenness centrality of node in the directed graph; is the total number of nodes in the directed graph; Step 2.4: After obtaining the standardized betweenness centrality of each node, select the top nodes with the largest values as key nodes to expand the connection relationship of the nodes with low degrees; the specific process of expanding the connection relationship is: Define Indicates the th node The current set of key nodes that can be considered for connection; represented by Indicates the th node The scoring function of. If the current node is a Xiaodu node, it will select a node with the optimal score from the set of key nodes through the scoring function If there is no current connection between the Xiaodu node and the node with the optimal score, a connection is established between the current Xiaodu node and the node with the optimal score.

[0009] Furthermore, the specific process of step 3 is as follows: Step 3.1: Define the meta-path set of the heterogeneous graph , is the th meta-path; Calculate the attention value of each node on each meta-path: (7); Among them, is the th node on the th meta-path The attention value of, ; is the th node on the th meta-path The neighbor node set of; Is the exponential function with base e; Is the activation function; Represents the connection operation; is the th meta-path The attention vector of; Is the feature vector of the th node after feature transformation; Is the feature vector of the th node after feature transformation. Here, the th node is in The th neighbor node; (8); Among them, is the th node on the a meta - path 's feature embedding vector; is an activation function; Step 3.3: After obtaining the feature embedding vectors of each node on each meta - path, stack the feature embedding vectors of all nodes on the same meta - path row - by - row to obtain a feature embedding matrix composed of all nodes on the same meta - path; then fuse the original node feature matrix with the feature embedding matrix: (9); where, is the th meta - path 's feature embedding matrix; is the th meta - path 's feature embedding matrix composed of all nodes; is the original node feature matrix; is a learnable parameter, representing the weight of the meta - path credibility, and dynamically adjusts the weight of each meta - path according to The calculation formula is: (10); where, is another learnable parameter; represents the sigmoid activation function.

[0010] Furthermore, the specific process of step 4 is as follows: Step 4.1: After obtaining the feature embedding matrices of different meta - paths after fusion, is the feature embedding matrix of the th meta - path after fusion, calculate the representation ability scores of different meta - paths: (11); where, is the representation ability score of the th meta - path ; is the set of all nodes in the heterogeneous graph; is the attention vector; and are different learnable parameters; is the th node in the th meta - path 's feature embedding vector; Step 4.2: Normalize the representation ability scores to obtain attention weights: (12); Among them, is the th meta - path 's attention weight; is the number of meta - paths; is the th meta - path 's representation ability score; Step 4.3: Obtain the final embedding matrix representation of the nodes according to the attention weights: (13); Among them, is the th node 's final embedding matrix representation; is the th meta - path 's attention weight.

[0011] Furthermore, in the said step 5, the entire model is optimized by minimizing the following loss function: (14); Among them, is the value of the loss function; is the set of nodes; is the - th order node in the node set ; is 's true label; is the classifier parameter; is 's final embedding matrix representation.

[0012] Furthermore, in the said step 6, for the user node, obtain the final embedding matrix representation of the user node according to formula (13), for the novel node, obtain the final embedding matrix representation of the novel node; perform similarity calculation, calculate the similarity between each node in the final embedding matrix representation of the user node and the final embedding matrix representation of the novel node; sort the similarity values in descending order, and select the top similarity values corresponding novels to generate a personalized recommendation list. At this time, the recommended list contains the novels that best match the user's preferences.

[0013] The beneficial technical effects brought by the present invention are as follows.

[0014] 1. Enhance the feature aggregation ability of Xiaodu nodes: By using the node importance discovery algorithm to identify the key nodes in the novel recommendation graph and connecting the Xiaodu nodes to these key nodes, HGNN-CERE can significantly improve the connectivity of Xiaodu nodes and enhance their feature aggregation ability. This strategy effectively overcomes the problem of scarce information that Xiaodu nodes may have in the recommendation system, improves the representation ability of low-degree nodes, and enables the recommendation model to more accurately capture users' potential interests in unpopular novels.

[0015] 2. Purify the influence of unreliable meta-paths: During the meta-path aggregation process, traditional methods may be affected by unreliable meta-paths, resulting in the model learning noisy information. The purification mechanism proposed by HGNN-CERE ensures more accurate and reliable information in the feature aggregation process by offsetting the influence of unreliable meta-paths, thereby improving the performance and stability of the recommendation system.

[0016] 3. Introduction of path-level aggregation strategy: Through the path-level aggregation strategy, HGNN-CERE can fuse the semantic representations from different paths, thereby generating richer and more comprehensive node feature embeddings. This enables the model to comprehensively consider various potential user interests and novel features, further improving the accuracy and personalization of recommendations.

[0017] 4. Better adaptation to the novel recommendation task: By combining the above technologies, HGNN-CERE can better capture the complex heterogeneous relationships in novel recommendations, optimizing the modeling and prediction processes of the novel recommendation system. This method enables the recommendation system to achieve efficient prediction in various scenarios and provide more personalized novel recommendations for users, improving the user experience and satisfaction.

[0018] The method of the present invention solves the problems of insufficient feature aggregation ability of Xiaodu nodes and interference from unreliable meta-paths, significantly improves the accuracy of novel recommendations and the scalability of the system, and has high practical application value and broad market prospects. Brief Description of the Drawings

[0019] Figure 1 It is a flowchart of the novel recommendation method based on the neural network platform with enhanced node features of the present invention.

[0020] Figure 2 It is a schematic diagram of the visualization result of the existing HAN model.

[0021] Figure 3 It is a schematic diagram of the visualization result of the existing MAGNN model.

[0022] Figure 4 It is a schematic diagram of the visualization result of the existing GTN model.

[0023] Figure 5Schematic diagram of the visualization result of the existing HPN model.

[0024] Figure 6 Schematic diagram of the visualization result of the existing HGNN-QSSA model.

[0025] Figure 7 Schematic diagram of the visualization result of the HGNN-CERE model proposed by the present invention. Detailed implementation manners

[0026] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners: The present invention proposes a neural network platform novel recommendation method based on node feature enhancement, and constructs a heterogeneous graph neural network model HGNN-CERE based on node feature enhancement, which is more in line with the user's preferences when the platform novel is personalized recommended to users. This model mainly includes two innovative strategies: First, in order to enhance the feature acquisition ability of the Xiaodu node, HGNN-CERE innovatively introduces a node importance discovery algorithm. This algorithm can identify key nodes in the novel recommendation graph, which play an important bridging role in the associations between users and novels, novels and platforms, and authors and novels, and carry the key functions of information dissemination and feature transfer. By connecting these key nodes with the Xiaodu node, HGNN-CERE constructs an efficient information sharing mechanism, enabling the Xiaodu node to aggregate more feature information through these bridging nodes. This method effectively makes up for the deficiency of the Xiaodu node in information aggregation, and at the same time promotes the global flow of information in the novel recommendation graph, thus providing a more comprehensive feature representation for each node.

[0027] In addition, in order to reduce the negative impact of unreliable meta-paths on the model performance, HGNN-CERE proposes a purification strategy. During the meta-path feature aggregation process of novel recommendation, this strategy dynamically adjusts the reliability of each meta-path by introducing learnable parameters. When a certain meta-path is determined to be unreliable, the value of this parameter will be automatically adjusted to enhance the weight of the node's own features, thereby weakening the influence of the unreliable meta-path. This method ensures that the model can more accurately aggregate and represent node features in a complex heterogeneous relationship network, and avoids the decline of the recommendation effect caused by meta-path noise.

[0028] As Figure 1 shown, a neural network platform novel recommendation method based on node feature enhancement constructs a heterogeneous graph neural network model HGNN-CERE based on node feature enhancement. This model includes a Xiaodu node connectivity enhancement module, a path purification module, and an inter-path aggregation module; the method specifically includes the following steps: Step 1: Collect all the historical data of all novels of all users from the current novel platform, process it into the form of a heterogeneous graph, and perform feature transformation. The specific process is as follows: Step 1.1: The nodes in the heterogeneous graph represent different entities. Since the historical data contains multiple entities such as novels, authors, and users, the nodes in the heterogeneous graph include multiple types such as novels, authors, users, and novel platforms. The edges of the heterogeneous graph represent the relationships between entities, specifically including reading relationships, creation relationships, and subordination relationships. A heterogeneous graph is a graph composed of multiple types of nodes. According to the historical data, a node feature matrix for each entity is constructed, and all the node feature matrices form the required heterogeneous graph. Among them, the node feature matrix includes a novel node feature matrix, an author node feature matrix, and a user node feature matrix. The novel node feature matrix is composed of all novel information, the author node feature matrix is composed of all author information, and the user node feature matrix is composed of user reading records.

[0029] Among the precursor nodes of a novel node, there are user nodes that read the novel or other novel nodes that recommend the novel. After identifying the types of precursor nodes, the novel data is divided into regions according to these types, and then the region set is determined.

[0030] Identify the neighbor nodes with different types from the target node type in each region. If the target node is a novel, then the author, category, and user are all heterogeneous neighbors of the target node. At this time, the author, category, and user are heterogeneous neighbors. Combine the heterogeneous neighbors of all regions to form the heterogeneous neighbor set of the target node.

[0031] Step 1.2: Since the original node features in the node feature matrix may be in different feature spaces, first perform feature transformation for a certain type of node: (1); Among them, is the original feature of the th node , is the original feature dimension of the node; is the transformation matrix of a certain type of node, is the transformed node feature dimension; is the feature vector of the th node after feature transformation.

[0032] Step 2: Use the Xiaodu node connectivity enhancement module to expand the connection relationship of Xiaodu nodes. The Xiaodu node connectivity enhancement module uses the node importance discovery algorithm to identify the key node information in the heterogeneous graph, and then connects the Xiaodu nodes with the key node information to improve the connectivity of the Xiaodu nodes, so as to achieve the purpose of enhancing the feature aggregation ability of the Xiaodu nodes. The specific process is as follows: Step 2.1: To enhance the connectivity of Xiaodu nodes, it is first necessary to identify the Xiaodu nodes in the graph. For each node in the heterogeneous graph, its degree value is first calculated: (2); Among them, represents the adjacency matrix of the heterogeneous graph ; is the adjacency matrix in the th row and the th column, indicating whether the th node and the th node are connected. The th row corresponds to the th node, and the th column corresponds to the th node; represents the total number of all nodes; is the degree value of node .

[0033] After obtaining the degree values of all nodes, for each node, its area is traversed, and HGNN-CERE uses a ranking function to select the Xiaodu nodes among them: (3); Among them, is the list of degree values of all nodes; is the ranking function, which is used to rank the degree values and, after passing through the selection function , selects the top nodes as the Xiaodu nodes; is the list of Xiaodu nodes after selection.

[0034] Step 2.2: To discover the key nodes in the graph, HGNN-CERE uses betweenness centrality as a measure to judge the importance of nodes.

[0035] For node , its betweenness centrality is defined as the proportion of the shortest paths between all node pairs that pass through node : (4); Among them, is the betweenness centrality of node ; is from the th node to the th node the number of paths passing through node in the shortest path; is from the th node to the th node the number of shortest paths;

[0036] Step 2.3. To make the betweenness centrality comparable between graphs of different sizes, the betweenness centrality is usually standardized. The heterogeneous graph contains directed graphs and undirected graphs. In the novel recommendation scenario, there are nodes such as user nodes and novel nodes, and the edge types include the reading relationship between users and novels, the co-reading relationship between users, etc. Among them, the relationship between users and novels is directed, forming a directed graph, while the relationship between users is undirected, forming an undirected graph. Through the standardization process, the key nodes in the undirected graph and the directed graph can be found respectively.

[0037] For an undirected graph, the maximum betweenness centrality occurs in a star network, and the betweenness centrality of the central node is , which is the maximum value. Therefore, the betweenness centrality can be divided by this maximum value to obtain the standardized betweenness centrality: (5); where is the standardized betweenness centrality of node in the undirected graph; is the total number of nodes in the undirected graph.

[0038] For a directed graph, since there may be shortest paths in two directions for each pair of nodes and , the maximum possible betweenness centrality is . The corresponding standardization formula is: (6); where is the standardized betweenness centrality of node in the directed graph; is the total number of nodes in the directed graph.

[0039] Step 2.4. After obtaining the standardized betweenness centrality of each node, HGNN-CERE selects the top A node is used as a key node to expand the connection relationship of the Xiaodu node. If the Xiaodu node is connected to all key nodes, it may bring additional noise to the Xiaodu node. Therefore, HGNN-CERE selects any key node for each Xiaodu node according to the connection relationship of the network and constructs a connection relationship between them. However, it should be noted that if there is already a key node among the direct neighbors of the Xiaodu node, it does not need to add an additional connection edge. Therefore, use to represent the th node the current set of key nodes that can be considered for connection, usually including key nodes. Use to represent the th node 's scoring function, which is used to evaluate a certain property (such as distance, similarity, etc.) of the th node . Specifically, if the current node is a Xiaodu node, it will select a node with the optimal score from the set of key nodes through the scoring function , and ensure that there is no current connection between the Xiaodu node and the node with the optimal score. If these conditions are met, a connection is established between the current Xiaodu node and the node with the optimal score.

[0040] Step 3. The meta-path in the heterogeneous graph can be understood as a specific interaction sequence, such as user-novel interaction, which contains specific user reading preferences and novel content characteristics. In order to avoid the influence of unreliable meta-paths during meta-path aggregation, the path purification module of HGNN-CERE uses a purification mechanism to offset the influence of unreliable meta-paths.

[0041] Since the meta-path contains specific semantic information, HGNN can capture the feature embedding containing a certain semantics through the meta-path. However, as mentioned above, there are unreliable meta-paths in the randomly selected meta-paths, and this kind of unreliable meta-path cannot be discovered in advance. Therefore, purifying the meta-path and timely reducing the proportion of the unreliable meta-path in the feature embedding is an important method to reduce unreliable information. The specific process is as follows: Step 3.1. Define the set of meta-paths of the heterogeneous graph, as the th meta-path; calculate the attention value of each node in each meta-path: (7); where, is the attention value of the th node in the th meta-path , ; is the th node in the th meta - path 's neighbor node set; is the exponential function with base e; is the activation function; represents the concatenation operation; is the th meta - path 's attention vector, used to calculate the attention value; is the th node 's feature vector after feature transformation; is the feature vector of the th node after feature transformation, where the th node is the th neighbor node in ; Step 3.2: Use the attention mechanism to perform in - meta - path aggregation to obtain the feature embedding vector of each node in each meta - path: (8); where, is the th node in the th meta - path 's feature embedding vector; is the activation function, here set as the Tanh activation function.

[0042] Step 3.3: After obtaining the feature embedding vectors of each node in each meta - path, stack the feature embedding vectors of all nodes in the same meta - path row - by - row to obtain the feature embedding matrix composed of all nodes in the same meta - path. In this way, HGNN - CERE retains the original node features. At the same time, to avoid the influence of unreliable meta - paths, HGNN - CERE uses an adjustable parameter to fuse the original features with the feature embedding matrix: (9); where, is the fused th meta - path 's feature embedding matrix; is the th meta - path 's feature embedding matrix composed of all nodes; is the original node feature matrix; is the learnable parameter, representing the weight of path credibility, and the calculation formula is: (10); Among them, is another learnable parameter; represents the sigmoid activation function.

[0043] When the th meta-path belongs to the reliable meta-path, the value increases, increasing the weight, and enhancing the meta-path feature. When the meta-path belongs to the unreliable meta-path, the value decreases, reducing the weight, and enhancing the original feature. Specifically: HGNN-CERE utilizes a learnable parameter maintained through the sigmoid activation function . Specifically, if the meta-path makes a positive contribution to the novel recommendation task objective (i.e., during the learning process in the experimental stage, if a certain meta-path improves the predicted result after learning, it is determined that this path makes a positive contribution), the model will learn to increase the value of , so that the weight of the meta-path credibility approaches 1, which is the reliable meta-path. If the meta-path makes no contribution to the novel recommendation task objective or introduces noise, the model will learn to decrease the value of , so that the weight of the meta-path credibility approaches 0, which is the unreliable meta-path. By this way of dynamically adjusting the weight of each meta-path with a learnable parameter to judge the reliability of the meta-path, the weight of the reliable meta-path can be enhanced, and the weight of the unreliable meta-path can be weakened, thereby optimizing the feature representation and improving the recommendation performance.

[0044] Step 4: The path-level aggregation module adopts a path-level aggregation strategy to fuse the feature embeddings of different meta-paths and generate the final embedding matrix representation of the nodes. The specific process is as follows: Step 4.1: After obtaining the fused feature embedding matrices of different meta-paths , is the feature embedding matrix of the th meta-path after fusion. Calculate the representation ability scores of different meta-paths: (11); Among them, is the representation ability score of the th meta-path ; is the set of all nodes in the heterogeneous graph; is the attention vector; and are different learnable parameters; is the th node in the th meta-path feature embedding vector; Step 4.2. Normalize the representation ability score to obtain the attention weight: (12); where is the th meta-path attention weight; is the number of meta-paths; is the th meta-path representation ability score; Step 4.3. After obtaining the weights, the final embedding matrix representation of the node can be obtained: (13); where is the th node final embedding matrix representation; is the th meta-path attention weight.

[0045] Step 5. Construct a loss function to continuously update the model parameters and perform backpropagation to optimize the training model HGNN-CERE.

[0046] After obtaining the final embedding of the node, optimize and train the entire model by minimizing the following loss function: (14); where is the loss function value; is the node set; is the th order node in the node set; is true label; is the classifier parameter; is final embedding matrix representation.

[0047] Step 6. Obtain the novel data of the current user, input it into the trained model HGNN-CERE, and generate a personalized novel recommendation list according to each user's preference.

[0048] For user nodes, the final embedding matrix representation of user nodes is obtained according to formula (13), and for novel nodes, the final embedding matrix representation of novel nodes is obtained; similarity calculation is performed to calculate the similarity between each node in the final embedding matrix representation of user nodes and the final embedding matrix representation of novel nodes; the similarity values ​​are sorted in descending order, and the top nodes are selected. A personalized recommendation list is generated for the novels corresponding to the similarity values, and the recommended list contains the novels that best match the user's preferences. The similarity calculation can adopt a common similarity calculation method.

[0049] In order to verify the feasibility and superiority of the HGNN-CERE model proposed in the present invention in the novel recommendation task, the present invention designed a comparative experiment and selected seven baseline methods, namely GCN, GAT, HAN, MAGNN, GTN, HPN, and HGNN-QSSA, for comparative analysis. GCN uses a standard graph convolutional network to learn node embedding of homogeneous graphs, which can capture local structural information, but cannot be directly applied to heterogeneous graphs. GAT assigns attention weights to different neighbors through a graph attention mechanism for feature aggregation. HAN converts heterogeneous graphs into isomorphic subgraphs through meta-paths, and aggregates node representations under different semantics using semantic attention. MAGNN uses intra-meta-path aggregation and inter-meta-path aggregation to learn node embedding representations of heterogeneous graphs. GTN generates neighborhood graphs of different meta-paths by multiplying adjacency matrices, thereby automatically discovering valuable meta-paths. HPN combines efficient path aggregation based on graph structure to mine potential semantic information in heterogeneous graphs. HGNN-QSSA further optimizes the contribution of meta-paths to node representation learning through specific path selection and attention mechanisms.

[0050] To ensure fair comparison, the embedding dimension and dropout value of all models are set to the same value, where the embedding dimension is set to 64 and the dropout value is set to 0.6. For models that require meta-paths, such as HAN, MAGNN, GTN, etc., the meta-path selected in the novel recommendation task is {UPU, UBU, UPCU}, where U, P, B, and C represent users, novels, authors, and novel platforms, respectively. For the HGNN-QSSA model, the initial features are processed using linear transformations as input to the model. For models that use attention mechanisms, such as GAT and HGNN-QSSA, the number of attention heads is set to 8.

[0051] In the experiment, and Represents the dimension of the attention vector (if any) and the number of layers of the model: For GCN and GAT, set ; For HAN and MAGNN, set , ; For GTN, set the batch size to 16 and the number of neighbor samplings to 100; for HPN, adopt the best parameters recommended in its paper; for HGNN-QSSA, set , .

[0052] Adopt a support vector machine (SVM) as the classifier to evaluate the performance of the model in multi-classification tasks. The evaluation metrics are Macro-F1 (Ma-F1) and Micro-F1 (Mi-F1). The experimental results are shown in Table 1, and the best results are shown in bold. Through comparative analysis, HGNN-CERE outperforms other baseline models in both metrics on all datasets, verifying the superiority of the present invention in the novel recommendation task.

[0053] Table 1 Classification results of different models on three datasets (%) .

[0054] As shown in Table 1, in the ACM dataset, both the Macro-F1 and Micro-F1 of HGNN-CERE are higher than those of other mainstream models, such as GCN and GAT. Although MAGNN and GTN perform well, they still fail to exceed HGNN-CERE. In the Yelp dataset, HGNN-CERE is also superior to most models, especially with a Micro-F1 score of 92.71, showing good classification performance. Although the performance in the DBLP dataset is slightly weaker, HGNN-CERE is still higher than most of the comparison models. This superiority is mainly due to the fact that HGNN-CERE can purify unreliable meta-paths, improve the aggregation ability between paths, and the enhancement of small-degree nodes also brings a certain degree of performance improvement.

[0055] Next, use the normalized mutual information (NMI) and adjusted rand index (ARI) as evaluation metrics to view the experimental performance of the model in clustering tasks. The experimental results are shown in Table 2.

[0056] As shown in Table 2, in the clustering experiment, the performance of the HGNN-CERE model shows significant advantages. In the ACM dataset, the NMI of HGNN-CERE reaches 0.7196, far exceeding other models, indicating that it has higher consistency and accuracy in clustering performance.

[0057] In the Yelp dataset, HGNN-CERE once again demonstrated superiority, with an NMI of 0.6758, higher than most baseline models. On the DBLP dataset, the NMI value of HGNN-CERE was 0.7622, and the ARI reached 0.6337. Although these two metrics were not as significant as those in the ACM dataset, they still proved the clustering effect of HGNN-CERE. This is mainly because after removing unreliable meta-paths, the feature aggregation ability of nodes is stronger.

[0058] Table 2 Results of different models in the clustering experiment (%) 。

[0059] To provide a more intuitive evaluation, t-SNE was used to project the node embeddings into a two-dimensional space and color them according to their labels. Figures 2 - 7 The visualization results of the node distributions of the existing HAN, MAGNN, GTN, HPN, HGNN-QSSA models and the HGNN-CERE model proposed in the present invention on the ACM dataset are respectively shown. The figure shows the distributions of three types of nodes (category 0, category 1, category 2). Compared with other models, the node type gaps of the HGNN-CERE model proposed in the present invention are larger, indicating that the node types of the model proposed in the present invention are more distinguishable.

[0060] The experimental results show that the application of the model proposed in the present invention in platform novel recommendation is effective and superior to existing recommendation methods. This achievement provides a new idea and method for the field of platform novel recommendation, which helps to better understand and predict users' novel preferences and behavior patterns.

[0061] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A novel recommendation method for a neural network platform based on enhanced node features, characterized in that, Construct a heterogeneous graph neural network model HGNN-CERE based on enhanced node features. This model includes a Xiaodu node connectivity enhancement module, a path purification module, and an inter-path aggregation module. The method specifically includes the following steps: Step 1: Collect all the historical novel data of all users from the current novel platform, process it into the form of a heterogeneous graph, and perform feature transformation. Step 2: The Xiaodu node connectivity enhancement module uses the node importance discovery algorithm to identify the key node information in the heterogeneous graph, and then connects the Xiaodu node with the key node information to expand the connection relationship of the Xiaodu node. Step 3: Dynamically adjust the weight of each meta-path based on the path purification module. Step 4: The inter-path aggregation module adopts a path-level aggregation strategy to fuse the feature embeddings of different meta-paths and generate the final embedding matrix representation of the nodes. Step 5: Construct a loss function to optimize and train the model HGNN-CERE. Step 6: Obtain the novel data of the current user, input it into the trained model HGNN-CERE, and generate a personalized novel recommendation list according to the preferences of each user.

2. The novel recommendation method of the neural network platform based on node feature enhancement according to claim 1, wherein, The specific process of Step 1 is as follows: Step 1.1: The nodes in the heterogeneous graph represent different entities. Since the historical data contains multiple entities such as novels, authors, and users, the nodes in the heterogeneous graph include multiple types such as novels, authors, users, and novel platforms. The edges of the heterogeneous graph represent the relationships between entities, specifically including reading relationships, creation relationships, and subordination relationships. The heterogeneous graph is a graph composed of multiple types of nodes. According to the historical data, the node feature matrix of each entity is constructed, and all the node feature matrices form the required heterogeneous graph. Among them, the node feature matrix includes a novel node feature matrix, an author node feature matrix, and a user node feature matrix. The novel node feature matrix is composed of all novel information, the author node feature matrix is composed of all author information, and the user node feature matrix is composed of user reading records. Step 1.2: The original node features in the node feature matrix are in different feature spaces, and feature transformation is performed on the nodes. (1)； Among them, is the original feature of the th node, and is the original feature dimension of the node; is the transformation matrix of a certain type of node, and is the feature dimension of the node after transformation; is the feature vector of the th node after feature transformation.

3. The novel recommendation method based on a neural network platform with enhanced node features according to claim 2, wherein The specific process of Step 2 is as follows: Step 2.1: Calculate the degree value of each node in the heterogeneous graph. (2)； Among them, represents the adjacency matrix of the heterogeneous graph; is the adjacency matrix in the row and column value, indicating whether the th node and the th node are connected. The th row corresponds to the th node, and the th column corresponds to the th node; represents the total number of all nodes; is the degree value of node . After obtaining the degree values of all nodes, for each node, traverse its area and use a ranking function to select the Xiaodu nodes among them. (3)； Among them, is the list of degree values of all nodes; is the ranking function, used to rank the degree values, and after passing through the selection function select the top nodes as the low-degree nodes; is the list of low-degree nodes after selection; Step 2.2: Use betweenness centrality as a measurement criterion to judge the importance of nodes. (4)； Among them, is the betweenness centrality of node ; is the number of paths passing through node in the shortest path from the -th node to the -th node ; is the number of shortest paths from the -th node to the -th node ; Step 2.3: Standardize the betweenness centrality. The heterogeneous graph contains both directed graphs and undirected graphs. The relationship between users and novels is directed, forming a directed graph, while the relationship between users and users is undirected, forming an undirected graph. For the undirected graph, the standardized betweenness centrality is: (5)； Among them, is the normalized betweenness centrality of nodes in an undirected graph ; is the total number of nodes in the undirected graph; For the directed graph, the standardized betweenness centrality is: (6)； Among them, is the normalized betweenness centrality of the node in the directed graph ; is the total number of nodes in the directed graph; Step 2.4, after obtaining the normalized betweenness centrality of each node, select the top nodes as key nodes to expand the connection relationship of the Xiaodu node; the specific process of expanding the connection relationship is as follows: Define to represent the th node the set of key nodes that can be considered for connection currently; use to represent the th node 's scoring function. If the current node is a Xiaodu node, it will select a node with the optimal score from the key node set through the scoring function . If there is no current connection between the Xiaodu node and the node with the optimal score, a connection is established between the current Xiaodu node and the node with the optimal score.

4. The novel recommendation method based on a neural network platform with enhanced node features according to claim 3, characterized in that The specific process of Step 3 is as follows: Step 3.1: Define the meta-path set of the heterogeneous graph , is the th meta-path; Calculate the attention value of each node on each meta-path: (7)； Among them, is the th node in the th meta - path 's attention value; ; is the th node in the th meta - path 's neighbor node set; is the exponential function with base e; is the activation function; represents the concatenation operation; is the th meta - path 's attention vector; is the feature vector of the th node after feature transformation; is the feature vector of the th node after feature transformation, where the th node is the th th neighbor node in Step 3.2: Adopt an attention mechanism to perform intra-path aggregation and obtain the feature embedding vector of each node in each meta-path. (8)； Among them, is the th node in the th meta-path feature embedding vector; is the activation function; Step 3.3: After obtaining the feature embedding vectors of each node on each meta-path, stack the feature embedding vectors of all nodes on the same meta-path row by row to obtain a feature embedding matrix composed of all nodes on the same meta-path; then fuse the original node feature matrix with the feature embedding matrix: (9)； Among them, is the feature embedding matrix of the th meta-path after fusion; is the feature embedding matrix composed of all nodes of the th meta-path ; is the original node feature matrix; is a learnable parameter representing the weight of the meta-path credibility, and dynamically adjusts the weight of each meta-path according to , and the calculation formula is: (10)； Among them, is another learnable parameter; represents the sigmoid activation function.

5. The novel recommendation method for the neural network platform based on node feature enhancement according to claim 4, wherein The specific process of step 4 is as follows: Step 4.

1. After obtaining the feature embedding matrix of different meta-paths after fusion then, For the feature embedding matrix of the th meta-path after fusion calculate the representation ability scores of different meta-paths: (11)； Among them, is the representation ability score of the th meta - path; is the set of all nodes in the heterogeneous graph; is the attention vector; and are different learnable parameters; is the th node in the th meta - path feature embedding vector; Step 4.2: Normalize the representation ability score to obtain the attention weights: (12)； Among them, is the th attention weight of the th meta-path; is the number of meta-paths; is the th representation ability score of the th meta-path; Step 4.3: Obtain the final embedding matrix representation of the nodes according to the attention weights: (13)； Among them, is the final embedding matrix representation of the th node; is the attention weight of the th meta-path.

6. The novel recommendation method based on a neural network platform with enhanced node features according to claim 5, wherein In step 5, the entire model is optimized by minimizing the following loss function: (14)； Among them, is the loss function value; is the node set; is the node set in the node of order; is the true label of; is the classifier parameter; is the final embedding matrix representation of.

7. The novel recommendation method of the neural network platform based on node feature enhancement according to claim 6, wherein In step 6, for user nodes, the final embedding matrix representation of user nodes is obtained according to formula (13), and for novel nodes, the final embedding matrix representation of novel nodes is obtained; Calculate the similarity, and calculate the similarity between each node in the final embedding matrix representation of the user node and the final embedding matrix representation of the novel node; sort the similarity values in descending order, and select the top novels corresponding to the similarity values to generate a personalized recommendation list. At this time, the recommendation list contains the novels that best match the user's preferences.

Images