A novel recommendation method for neural network platform based on node feature enhancement

Abstract

The present invention discloses a novel recommendation method for a neural network platform based on node feature enhancement, which belongs to the field of deep learning. It constructs a heterogeneous graph neural network model based on node feature enhancement, which includes a small-degree node connectivity enhancement module, a path purification module, and an inter-path aggregation module. The method specifically includes the following steps: collecting all novel history data of all users from the current novel platform, processing them into a heterogeneous graph form, and performing feature conversion; expanding the connection relationship of small-degree nodes based on the small-degree node connectivity enhancement module; purifying the meta-path based on the path purification module; generating the final embedding matrix representation of the node based on the inter-path aggregation module; constructing a loss function optimization training model; obtaining the novel data of the current user, inputting the trained model, and generating a personalized novel recommendation list according to the preferences of each user. The present invention significantly improves the accuracy of novel recommendations and recommends novels that best match the user's preferences.

Description

Technical Field

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

[0002] Currently, there are several problems when recommending novels based on traditional heterogeneous graph neural networks.

[0003] First, there's the performance issue of low-degree nodes. Heterogeneous graph neural networks contain multiple types of nodes and edges. In the novel recommendation scenario, nodes typically include users, novels, authors, and novel platforms, while edges represent interactions between users and novels (e.g., reading, liking, and saving), as well as associations between novels and categories and authors. Some nodes, due to their limited connections to other nodes (i.e., low degree), may be unable to effectively aggregate sufficient feature information from their neighboring nodes. For example, some user nodes may have only interacted with a small number of novel nodes, resulting in insufficient feature information about the novels. Similarly, some unpopular novel nodes, due to their limited interactions with users, also lack sufficient information about user preferences. In contrast, user nodes with higher degrees or popular novel nodes can acquire more neighboring feature information through rich interactions, resulting in more expressive representations. This lack of information aggregation limits the representational power of low-degree nodes, resulting in poor performance in subsequent novel recommendation tasks. This not only impacts the performance of individual nodes but can also indirectly affect feature learning across the entire graph due to limited information transfer, ultimately reducing the performance of the overall recommendation model.

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

[0005] In order to solve the above problems, the present invention proposes a novel recommendation method for a neural network platform based on node feature enhancement, and constructs a heterogeneous graph neural network model HGNN-CERE based on node feature enhancement. In order to improve the feature aggregation ability of small-degree nodes, HGNN-CERE first uses the node importance algorithm to identify important node information in the graph, and then connects the small-degree nodes with these important nodes to improve the connectivity of the small-degree nodes, so as to achieve the purpose of enhancing the feature aggregation ability of the small-degree nodes. 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 solutions of the present invention are as follows:

[0007] A novel recommendation method for a neural network platform based on node feature enhancement is proposed. A heterogeneous graph neural network model HGNN-CERE based on node feature enhancement is constructed. The model includes a small-degree node connectivity enhancement module, a path purification module, and an inter-path aggregation module. The method specifically includes the following steps:

[0008] Step 1: Collect all historical novel data of all users from the current novel platform, process it into a heterogeneous graph, and perform feature conversion;

[0009] Step 2: The small degree node connectivity enhancement module uses the node importance discovery algorithm to identify the key node information in the heterogeneous graph, and then connects the small degree nodes with the key node information to expand the connection relationship of the small degree nodes;

[0010] Step 3: Dynamically adjust the weight of each meta-path based on the path purification module;

[0011] 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 node;

[0012] Step 5: Construct the loss function optimization training model HGNN-CERE;

[0013] Step 6: Obtain the novel data of the current user, input the trained model HGNN-CERE, and generate a personalized novel recommendation list based on each user's preferences.

[0014] Furthermore, the specific process of step 1 is as follows:

[0015] Step 1.1. The nodes in the heterogeneous graph represent different entities. Since historical data contains multiple entities such as novels, authors, and users, the nodes in the heterogeneous graph include novels, authors, users, and novel platforms. The edges of the heterogeneous graph represent the relationships between entities, including reading relationships, creation relationships, and subordinate relationships. A heterogeneous graph is composed of multiple types of nodes. Based on historical data, a node feature matrix is ​​constructed for each entity, and all node feature matrices constitute the required heterogeneous graph. 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 consists of all novel information, the author node feature matrix consists of all author information, and the user node feature matrix consists of user reading records.

[0016] Step 1.2: The original node features in the node feature matrix are located in different feature spaces, and the node features are transformed:

[0017] (1);

[0018] in, It is nodes The original characteristics of 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; After feature transformation nodes The eigenvector of .

[0019] Furthermore, the specific process of step 2 is as follows:

[0020] Step 2.1. Calculate the degree of each node in the heterogeneous graph:

[0021] (2);

[0022] in, Representing the adjacency matrix of a heterogeneous graph; is the adjacency matrix Middle Rank The value of the column indicates nodes Hedi nodes Is it connected? The row corresponds to Node, Column corresponds to nodes; Indicates the number of all nodes; For nodes The degree value;

[0023] After obtaining the degree values ​​of all nodes, for each node, traverse its area and use a ranking function to select the nodes with small degrees:

[0024] (3);

[0025] in, is a list of degree values ​​of all nodes; Is a ranking function, used to rank the degree value, and after the selection function Select the front Nodes are regarded as small-degree nodes; is the list of small degree nodes after selection;

[0026] Step 2.2: Use betweenness centrality as a metric to determine the importance of a node:

[0027] (4);

[0028] in, is a node Betweenness centrality; It is from nodes To nodes The shortest path passes through the node The number of paths; It is from nodes To nodes The number of shortest paths;

[0029] Step 2.3: Standardize the betweenness centrality. The heterogeneous graph contains directed and undirected graphs. The relationship between users and novels is directed, forming a directed graph, while the relationship between users is undirected, forming an undirected graph. For the undirected graph, the standardized betweenness centrality is:

[0030] (5);

[0031] in, is a node in an undirected graph Normalized betweenness centrality; is the total number of nodes in the undirected graph;

[0032] For a directed graph, the normalized betweenness centrality is:

[0033] (6);

[0034] in, is a node in a directed graph Normalized betweenness centrality; is the total number of nodes in the directed graph;

[0035] Step 2.4: After obtaining the standardized betweenness centrality of each node, select the largest one. Nodes are used as key nodes to expand the connection relationship of small degree nodes; the specific process of expanding the connection relationship is: define Indicates the nodes The set of key nodes that can be considered for connection at present; Indicates the nodes If the current node is a small node, it will be selected from the key node set. Through the scoring function Select a node with the best score. If there is no connection between the small node and the node with the best score, establish a connection between the current small node and the node with the best score.

[0036] Furthermore, the specific process of step 3 is as follows:

[0037] Step 3.1: Define the meta-path set of the heterogeneous graph , For the meta-paths; calculate the attention value of each node in each meta-path:

[0038] (7);

[0039] in, It is nodes In the Metapath The attention value, ; It is nodes In the Metapath The set of neighbor nodes of is an exponential function with base e; is the activation function; Represents a connection operation; It is Metapath The attention vector of After feature transformation The feature vector of each node; After feature transformation The feature vector of the node, here The nodes are The neighbor nodes;

[0040] Step 3.2: Use the attention mechanism to perform meta-path aggregation and obtain the feature embedding vector of each node in each meta-path:

[0041] (8);

[0042] in, It is nodes In the Metapath The feature embedding vector of is the activation function;

[0043] Step 3.3: After obtaining the feature embedding vector of each node in each meta-path, stack the feature embedding vectors of all nodes in the same meta-path in rows to obtain the feature embedding matrix consisting of all nodes in the same meta-path; then fuse the original node feature matrix with the feature embedding matrix:

[0044] (9);

[0045] in, After fusion Metapath The feature embedding matrix of For the Metapath The feature embedding matrix composed of all nodes; is the original node feature matrix; is a learnable parameter that represents the weight of the meta-path credibility, according to Dynamically adjust the weight of each meta-path, the calculation formula is:

[0046] (10);

[0047] in, is another learnable parameter; Represents the sigmoid activation function.

[0048] Furthermore, the specific process of step 4 is as follows:

[0049] Step 4.1: Obtain the feature embedding matrix of different meta-paths after fusion back, After fusion Metapath The feature embedding matrix of , calculates the representation ability score of different meta-paths:

[0050] (11);

[0051] in, It is Metapath The expression ability score of is the set of all nodes in the heterogeneous graph; is the attention vector; and are different learnable parameters; It is nodes In the Metapath The feature embedding vector of

[0052] Step 4.2: Normalize the representation ability score to obtain the attention weight:

[0053] (12);

[0054] in, It is Metapath The attention weight of is the number of meta-paths; It is Metapath The expression ability score of

[0055] Step 4.3. According to the attention weight, obtain the final embedding matrix representation of the node:

[0056] (13);

[0057] in, It is nodes The final embedding matrix representation of ; It is Metapath The attention weight of .

[0058] Furthermore, in step 5, the entire model is optimized and trained by minimizing the following loss function:

[0059] (14);

[0060] in, is the loss function value; is a set of nodes; A node collection middle Node of order; for The true label of is the classifier parameter; for The final embedding matrix representation of .

[0061] Furthermore, in step 6, for the user node, the final embedding matrix representation of the user node is obtained according to formula (13), and for the novel node, the final embedding matrix representation of the novel node is obtained; similarity calculation is performed to 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; 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. At this time, the recommendation list contains the novels that best match the user's preferences.

[0062] The beneficial technical effects brought about by the present invention are as follows.

[0063] 1. Enhanced feature aggregation capabilities for small-degree nodes: By using a node importance discovery algorithm to identify key nodes in the novel recommendation graph and connecting small-degree nodes to these key nodes, HGNN-CERE significantly improves the connectivity of small-degree nodes and enhances their feature aggregation capabilities. This strategy effectively overcomes the information scarcity problem that small-degree nodes may present in recommendation systems, improves the representation capabilities of low-degree nodes, and enables the recommendation model to more accurately capture users' potential interest in less popular novels.

[0064] 2. Purifying the Impact of Unreliable Meta-Paths: Traditional methods can be affected by unreliable meta-paths during meta-path aggregation, causing the model to learn noisy information. The purification mechanism proposed by HGNN-CERE offsets the impact of unreliable meta-paths, ensuring more accurate and reliable information during feature aggregation, thereby improving the performance and stability of the recommendation system.

[0065] 3. Introduction of a path-level aggregation strategy: Through this path-level aggregation strategy, HGNN-CERE is able to fuse semantic representations from different paths to generate richer and more comprehensive node feature embeddings. This allows the model to comprehensively consider a variety of potential user interests and novel features, further improving the accuracy and personalization of recommendations.

[0066] 4. Better Adaptation to Novel Recommendation Tasks: By combining the aforementioned technologies, HGNN-CERE can better capture the complex heterogeneous relationships in novel recommendations, optimizing the modeling and prediction process of novel recommendation systems. This approach enables the recommendation system to achieve efficient predictions in a variety of scenarios and provide users with more personalized novel recommendations, improving user experience and satisfaction.

[0067] The method of the present invention solves the problems of insufficient feature aggregation capabilities of small-degree 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

[0068] Figure 1 This is a flow chart of the novel recommendation method based on the neural network platform with enhanced node features according to the present invention.

[0069] Figure 2 Schematic diagram for visualizing the results of the existing HAN model.

[0070] Figure 3 Schematic diagram for visualizing the results of the existing MAGNN model.

[0071] Figure 4 Schematic diagram for visualizing the results of existing GTN models.

[0072] Figure 5 Schematic diagram for visualizing the results of existing HPN models.

[0073] Figure 6 Schematic diagram for visualizing the results of the existing HGNN-QSSA model.

[0074] Figure 7 Schematic diagram of the visualization results of the HGNN-CERE model proposed in this invention. DETAILED DESCRIPTION

[0075] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0076] This paper proposes a novel recommendation method for online platforms based on neural network node feature enhancement. It constructs a heterogeneous graph neural network model, HGNN-CERE, based on node feature enhancement. This model is used to make personalized recommendations for online novels that better suit users' preferences. The model mainly includes two innovative strategies:

[0077] First, to enhance the feature acquisition capabilities of Xiaodu nodes, HGNN-CERE innovatively introduces a node importance discovery algorithm. This algorithm identifies key nodes in the novel recommendation graph. These nodes serve as important bridges between users and novels, novels and platforms, and authors and novels, carrying the key functions of information dissemination and feature transfer. By connecting these key nodes with Xiaodu nodes, HGNN-CERE constructs an efficient information sharing mechanism, enabling Xiaodu nodes to aggregate more feature information through these bridge nodes. This method effectively compensates for the shortcomings of Xiaodu nodes in information aggregation while promoting the global flow of information in the novel recommendation graph, thereby providing a more comprehensive feature representation for each node.

[0078] Furthermore, to mitigate the negative impact of unreliable meta-paths on model performance, HGNN-CERE proposes a purification strategy. During the meta-path feature aggregation process for novel recommendations, this strategy dynamically adjusts the reliability of each meta-path by introducing a learnable parameter. When a meta-path is deemed unreliable, the value of this parameter is automatically adjusted to increase the weight of the node's own features, thereby weakening the impact of the unreliable meta-path. This approach ensures that the model can more accurately aggregate and represent node features in complex, heterogeneous relational networks, preventing degradation of recommendation effectiveness due to meta-path noise.

[0079] like Figure 1 As shown in the figure, a novel recommendation method for a neural network platform based on node feature enhancement is proposed. A heterogeneous graph neural network model HGNN-CERE based on node feature enhancement is constructed. The model includes a small-degree node connectivity enhancement module, a path purification module, and an inter-path aggregation module. The method specifically includes the following steps:

[0080] Step 1: Collect all historical novel data of all users from the current novel platform, process it into a heterogeneous graph, and perform feature conversion; the specific process is as follows:

[0081] Step 1.1: Nodes in a heterogeneous graph represent different entities. Since historical data includes multiple entities such as novels, authors, and users, the nodes in the heterogeneous graph include novels, authors, users, and novel platforms. The edges of the heterogeneous graph represent relationships between entities, specifically reading relationships, creation relationships, and subordinate relationships. A heterogeneous graph is composed of multiple types of nodes. A node feature matrix for each entity is constructed based on historical data, and all node feature matrices together form the required heterogeneous graph. 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 consists of all novel information, the author node feature matrix consists of all author information, and the user node feature matrix consists of user reading records.

[0082] The predecessor nodes of a novel node include user nodes that read the novel or other novel nodes that recommend the novel. After identifying the types of the predecessor nodes, the novel data is divided into regions according to these types, and then the region sets are determined.

[0083] Identify neighbor nodes in each region that are different from the target node type. If the target node is a novel, then the author, category, and user are all heterogeneous neighbors of the target node. In this case, the author, category, and user are heterogeneous neighbors. Combine the heterogeneous neighbors of all regions to form a heterogeneous neighbor set of the target node.

[0084] Step 1.2: Since the original node features in the node feature matrix may be located in different feature spaces, feature transformation for a certain type of node is first performed:

[0085] (1);

[0086] in, It is nodes The original characteristics of 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; After feature transformation nodes The eigenvector of .

[0087] Step 2: Use the small degree node connectivity enhancement module to expand the connection relationship of small degree nodes. The small degree node connectivity enhancement module uses the node importance discovery algorithm to identify key node information in the heterogeneous graph, and then connects the small degree nodes with the key node information to improve the connectivity of the small degree nodes, so as to achieve the purpose of enhancing the small degree node feature aggregation capability. The specific process is as follows:

[0088] Step 2.1: To enhance the connectivity of small-degree nodes, we first need to identify the small-degree nodes in the graph. For each node in the heterogeneous graph, we first calculate its degree:

[0089] (2);

[0090] in, Representing heterogeneous graphs The adjacency matrix of is the adjacency matrix Middle Rank The value of the column indicates nodes Hedi nodes Is it connected? The row corresponds to Node, Column corresponds to nodes; Indicates the number of all nodes; For nodes The degree value.

[0091] After obtaining the degree values ​​of all nodes, HGNN-CERE traverses the area where each node is located and uses a ranking function to select the nodes with small degrees:

[0092] (3);

[0093] in, is a list of degree values ​​of all nodes; Is a ranking function, used to rank the degree value, and after the selection function Select the front Nodes are regarded as small-degree nodes; is the list of small degree nodes after selection.

[0094] Step 2.2: In order to discover the key nodes in the graph, HGNN-CERE uses betweenness centrality as a metric to judge the importance of the nodes.

[0095] For nodes , whose betweenness centrality is defined as the number of nodes in the shortest path between all pairs of nodes passing through the node Ratio:

[0096] (4);

[0097] in, is a node Betweenness centrality; It is from nodes To nodes The shortest path passes through the node The number of paths; It is from nodes To nodes The number of shortest paths.

[0098] Step 2.3: To make betweenness centrality comparable across graphs of different sizes, betweenness centrality is typically normalized. Heterogeneous graphs include both directed and undirected graphs. In the novel recommendation scenario, there are nodes such as user nodes and novel nodes. Edge types include reading relationships between users and novels, and shared reading relationships between users. The relationship between users and novels is directed, forming a directed graph, while the relationship between users is undirected, forming an undirected graph. Through normalization, key nodes can be identified in both undirected and directed graphs.

[0099] For undirected graphs, the maximum betweenness centrality occurs in star networks, where the betweenness centrality of the central node is , Therefore, we can divide the betweenness centrality by this maximum value to get the standardized betweenness centrality:

[0100] (5);

[0101] in, is a node in an undirected graph Normalized betweenness centrality; is the total number of nodes in the undirected graph.

[0102] For a directed graph, since each pair of nodes and There are possible shortest paths in both directions, so the maximum possible betweenness centrality is The corresponding normalization formula is:

[0103] (6);

[0104] in, is a node in a directed graph Normalized betweenness centrality; is the total number of nodes in the directed graph.

[0105] Step 2.4: After obtaining the normalized betweenness centrality of each node, HGNN-CERE selects the largest front Nodes are used as key nodes to expand the connection relationship of small degree nodes. If the small degree nodes are connected to all key nodes, it may bring additional noise to the small degree nodes. Therefore, HGNN-CERE selects any key node for each small degree node according to the connection relationship of the network and builds a connection relationship between them. However, it should be noted that if there is a key node in the direct neighbor of the small degree node, it does not need to add additional connection edges. Therefore, use Indicates the nodes The set of key nodes that can be considered for connection at present, usually including key nodes. Indicates the nodes The scoring function is used to evaluate the nodes Specifically, if the current node is a small node, it will be removed from the key node set. Through the scoring function Select a node with the best score and make sure there is no connection between the node with the smallest degree and the node with the best score. If these conditions are met, establish a connection between the current node with the smallest degree and the node with the best score.

[0106] Step 3: Metapaths in a heterogeneous graph can be understood as specific interaction sequences, such as user-novel interactions, which embody specific user reading preferences and novel content characteristics. To avoid the influence of unreliable metapaths during metapath aggregation, the path purification module of HGNN-CERE uses a purification mechanism to offset the impact of unreliable metapaths.

[0107] Since meta-paths contain specific semantic information, meta-path HGNN can capture feature embeddings containing certain semantics. However, as mentioned earlier, there are unreliable meta-paths in randomly selected meta-paths, and such unreliable meta-paths cannot be discovered in advance. Therefore, purifying meta-paths and promptly reducing the proportion of unreliable meta-paths in feature embeddings is an important method to reduce unreliable information. The specific process is as follows:

[0108] Step 3.1: Define the meta-path set of the heterogeneous graph , For the meta-paths; calculate the attention value of each node in each meta-path:

[0109] (7);

[0110] in, It is nodes In the Metapath The attention value, ; It is nodes In the Metapath The set of neighbor nodes of is an exponential function with base e; is the activation function; Represents a connection operation; It is Metapath The attention vector is used to calculate the attention value; After feature transformation nodes The eigenvector of After feature transformation The feature vector of the node, here The nodes are The neighbor nodes;

[0111] Step 3.2: Use the attention mechanism to perform meta-path aggregation and obtain the feature embedding vector of each node in each meta-path:

[0112] (8);

[0113] in, It is nodes In the Metapath The feature embedding vector of Is the activation function, which is set as Tanh activation function here.

[0114] Step 3.3: After obtaining the feature embedding vector of each node in each meta-path, the feature embedding vectors of all nodes in the same meta-path are stacked row by row to obtain the feature embedding matrix consisting 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:

[0115] (9);

[0116] in, After fusion Metapath The feature embedding matrix of For the Metapath The feature embedding matrix composed of all nodes; is the original node feature matrix; Is a learnable parameter that represents the weight of the path credibility, and the calculation formula is:

[0117] (10);

[0118] in, is another learnable parameter; Represents the sigmoid activation function.

[0119] When Metapath When it belongs to a reliable meta-path, The value increases, the weight increases, and the meta-path feature is enhanced. When it belongs to an unreliable meta-path, The value is reduced, the weight is reduced, and the original features are enhanced. Specifically: HGNN-CERE uses a learnable parameter After the sigmoid activation function, Specifically, if the meta-path It has a positive contribution to the novel recommendation task goal (that is, during the learning process of the experimental stage, if a meta-path improves the prediction results after learning, then the path is judged to have a positive contribution). The value of increases, so the weight of the meta-path credibility Close to 1, this is a reliable meta-path. It does not contribute to the novel recommendation task or introduces noise. The model will make The value of decreases, so the weight of the meta-path credibility If the weight is close to 0, it is an unreliable meta-path. By using a learnable parameter to dynamically adjust the weight of each meta-path to determine its reliability, the weight of reliable meta-paths can be increased and the weight of unreliable meta-paths can be decreased, thereby optimizing feature representation and improving recommendation performance.

[0120] Step 4: The inter-path aggregation module uses a path-level aggregation strategy to fuse the feature embeddings of different meta-paths and generate the final embedding matrix representation of the node. The specific process is as follows:

[0121] Step 4.1: Obtain the feature embedding matrix of different meta-paths after fusion back, After fusion Metapath The feature embedding matrix of , calculates the representation ability score of different meta-paths:

[0122] (11);

[0123] in, It is Metapath The expression ability score of is the set of all nodes in the heterogeneous graph; is the attention vector; and are different learnable parameters; It is nodes In the Metapath The feature embedding vector of

[0124] Step 4.2: Normalize the representation ability score to obtain the attention weight:

[0125] (12);

[0126] in, It is Metapath The attention weight of is the number of meta-paths; It is Metapath The expression ability score of

[0127] Step 4.3. After obtaining the weights, we can obtain the final embedding matrix representation of the node:

[0128] (13);

[0129] in, It is nodes The final embedding matrix representation of ; It is Metapath The attention weight of .

[0130] Step 5: Construct a loss function to continuously update and backpropagate model parameters to optimize the training model HGNN-CERE.

[0131] After obtaining the final embedding of the nodes, the entire model is optimized and trained by minimizing the following loss function:

[0132] (14);

[0133] in, is the loss function value; is a set of nodes; A node collection middle Node of order; for The true label of is the classifier parameter; for The final embedding matrix representation of .

[0134] Step 6: Obtain the novel data of the current user, input the trained model HGNN-CERE, and generate a personalized novel recommendation list based on each user's preferences.

[0135] 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. In this case, the recommendation list contains the novels that best match the user's preferences. The similarity calculation can adopt a common similarity calculation method.

[0136] To verify the feasibility and superiority of the proposed HGNN-CERE model in the novel recommendation task, we designed a comparative experiment and selected seven baseline methods: GCN, GAT, HAN, MAGNN, GTN, HPN, and HGNN-QSSA for comparative analysis. GCN uses a standard graph convolutional network to learn node embeddings in homogeneous graphs, which can capture local structural information but cannot be directly applied to heterogeneous graphs. GAT uses a graph attention mechanism to assign attention weights to different neighbors for feature aggregation. HAN converts heterogeneous graphs into isomorphic subgraphs using 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 in heterogeneous graphs. GTN generates neighborhood graphs for different meta-paths by multiplying adjacency matrices, thereby automatically discovering valuable meta-paths. HPN combines efficient path aggregation based on graph structure to mine latent 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.

[0137] To ensure fair comparison, all models were set to the same embedding dimension and dropout value: the embedding dimension was set to 64, and the dropout value was set to 0.6. For models that require a meta-path, such as HAN, MAGNN, and GTN, the meta-path used in the novel recommendation task is {UPU, UBU, UPCU}, where U, P, B, and C represent the user, novel, author, and platform of the novel, respectively. For the HGNN-QSSA model, the initial features are processed using linear transformations and used as model input. For models that use an attention mechanism, such as GAT and HGNN-QSSA, the number of attention heads is set to 8.

[0138] 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, the batch size is set to 16 and the number of neighbor samples is set to 100. For HPN, the best parameters recommended in the paper are adopted. For HGNN-QSSA, all datasets are set to , .

[0139] A support vector machine (SVM) was used as the classifier to evaluate the model's performance on a multi-classification task, using Macro-F1 (Ma-F1) and Micro-F1 (Mi-F1) as evaluation metrics. The experimental results are shown in Table 1, with the best results in bold. Comparative analysis shows that HGNN-CERE outperforms other baseline models in both metrics across all datasets, validating the superiority of this method in the novel recommendation task.

[0140] Table 1 Classification results of different models on three datasets (%)

[0141] .

[0142] As shown in Table 1, on the ACM dataset, HGNN-CERE's Macro-F1 and Micro-F1 scores are higher than those of other mainstream models, such as GCN and GAT. While MAGNN and GTN perform well, they still fail to surpass HGNN-CERE. On the Yelp dataset, HGNN-CERE also outperforms most models, particularly with a Micro-F1 score of 92.71, demonstrating excellent classification results. While its performance on the DBLP dataset is slightly weaker, HGNN-CERE still outperforms most of the comparison models. This superiority is primarily due to HGNN-CERE's ability to purify unreliable meta-paths, improving the ability to aggregate paths. Furthermore, the enhancement of low-degree nodes also contributes to a certain degree of performance improvement.

[0143] Next, we use normalized mutual information (NMI) and adjusted ranking index (ARI) as evaluation indicators to check the experimental performance of the model on the clustering task. The experimental results are shown in Table 2.

[0144] 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 effect.

[0145] On the Yelp dataset, HGNN-CERE again demonstrated superiority, achieving an NMI of 0.6758, surpassing most baseline models. On the DBLP dataset, HGNN-CERE achieved an NMI of 0.7622 and an ARI of 0.6337. While these two metrics are not as significant as those for ACM in this dataset, they still demonstrate HGNN-CERE's clustering effectiveness. This is primarily due to the improved feature aggregation capabilities of nodes after removing unreliable meta-paths.

[0146] Table 2 Results of different models in clustering experiments (%)

[0147] .

[0148] To provide a more intuitive evaluation, t-SNE is used to project the node embeddings into a two-dimensional space and color them according to their labels. Figure 2-Figure 7 The visualization results of the node distribution on the ACM dataset for the existing HAN, MAGNN, GTN, HPN, HGNN-QSSA models, and the proposed HGNN-CERE model are shown. The figure shows the distribution of three types of nodes (class 0, class 1, and class 2). Compared to the other models, the node type gaps in the proposed HGNN-CERE model are larger, indicating that the proposed model has a clearer distinction between node types.

[0149] Experimental results demonstrate that the proposed model is effective in platform novel recommendation and outperforms existing recommendation methods. This achievement provides a new approach and method for platform novel recommendation, helping to better understand and predict users' novel preferences and behavior patterns.

[0150] 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 technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A novel recommendation method based on neural network platform based on node feature enhancement, characterized in that: A heterogeneous graph neural network model HGNN-CERE based on node feature enhancement is constructed. The model includes a small-degree 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 historical novel data of all users from the current novel platform, process it into a heterogeneous graph, and perform feature conversion; Step 2: The small degree node connectivity enhancement module uses the node importance discovery algorithm to identify the key node information in the heterogeneous graph, and then connects the small degree nodes with the key node information to expand the connection relationship of the small degree nodes; Step 3: Dynamically adjust the weight of each meta-path based on the path purification module; the specific process is as follows: Step 3.1: Define the meta-path set of the heterogeneous graph , For the meta-paths; calculate the attention value of each node in each meta-path: (7)； in, It is nodes In the Metapath The attention value, ; It is nodes In the Metapath The set of neighbor nodes of is an exponential function with base e; is the activation function; Represents a connection operation; It is Metapath The attention vector of After feature transformation The feature vector of each node; After feature transformation The feature vector of the node, here The nodes are The neighbor nodes; Step 3.2: Use the attention mechanism to perform meta-path aggregation and obtain the feature embedding vector of each node in each meta-path: (8)； in, It is nodes In the Metapath The feature embedding vector of is the activation function; Step 3.3: After obtaining the feature embedding vector of each node in each meta-path, stack the feature embedding vectors of all nodes in the same meta-path in rows to obtain the feature embedding matrix consisting of all nodes in the same meta-path; then fuse the original node feature matrix with the feature embedding matrix: (9)； in, After fusion Metapath The feature embedding matrix of For the Metapath The feature embedding matrix composed of all nodes; is the original node feature matrix; is a learnable parameter that represents the weight of the meta-path credibility, according to Dynamically adjust the weight of each meta-path, the calculation formula is: (10)； in, is another learnable parameter; Represents the sigmoid activation function; 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 node; Step 5: Construct the loss function optimization training model HGNN-CERE; Step 6: Obtain the novel data of the current user, input the trained model HGNN-CERE, and generate a personalized novel recommendation list based on each user's preferences.

2. The novel recommendation method based on a neural network platform and node feature enhancement according to claim 1 is characterized in that: The specific process of step 1 is: Step 1.

1. The nodes in the heterogeneous graph represent different entities. Since historical data contains multiple entities such as novels, authors, and users, the nodes in the heterogeneous graph include novels, authors, users, and novel platforms. The edges of the heterogeneous graph represent the relationships between entities, including reading relationships, creation relationships, and subordinate relationships. A heterogeneous graph is composed of multiple types of nodes. Based on historical data, a node feature matrix is ​​constructed for each entity, and all node feature matrices constitute the required heterogeneous graph. 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 consists of all novel information, the author node feature matrix consists of all author information, and the user node feature matrix consists of user reading records. Step 1.2: The original node features in the node feature matrix are located in different feature spaces, and the node features are transformed: (1)； in, It is nodes The original characteristics of 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; After feature transformation nodes The eigenvector of .

3. The novel recommendation method based on a neural network platform and node feature enhancement according to claim 2 is characterized in that: The specific process of step 2 is: Step 2.

1. Calculate the degree of each node in the heterogeneous graph: (2)； in, Representing the adjacency matrix of a heterogeneous graph; is the adjacency matrix Middle Rank The value of the column indicates nodes Hedi nodes Is it connected? The row corresponds to Node, Column corresponds to nodes; Indicates the number of all nodes; For nodes The degree value; After obtaining the degree values ​​of all nodes, for each node, traverse its area and use a ranking function to select the nodes with small degrees: (3)； in, is a list of degree values ​​of all nodes; Is a ranking function, used to rank the degree value, and after the selection function Select the front Nodes are regarded as small-degree nodes; is the list of small degree nodes after selection; Step 2.2: Use betweenness centrality as a metric to determine the importance of a node: (4)； in, is a node Betweenness centrality; It is from nodes To nodes The shortest path passes through the node The number of paths; It is from nodes To nodes The number of shortest paths; Step 2.3: Standardize the betweenness centrality. The heterogeneous graph contains directed and undirected graphs. The relationship between users and novels is directed, forming a directed graph, while the relationship between users is undirected, forming an undirected graph. For the undirected graph, the standardized betweenness centrality is: (5)； in, is a node in an undirected graph Normalized betweenness centrality; is the total number of nodes in the undirected graph; For a directed graph, the normalized betweenness centrality is: (6)； in, is a node in a directed graph Normalized betweenness centrality; is the total number of nodes in the directed graph; Step 2.4: After obtaining the standardized betweenness centrality of each node, select the largest one. Nodes are used as key nodes to expand the connection relationship of small degree nodes; the specific process of expanding the connection relationship is: define Indicates the nodes The set of key nodes that can be considered for connection at present; Indicates the nodes If the current node is a small node, it will be selected from the key node set. Through the scoring function Select a node with the best score. If there is no connection between the small node and the node with the best score, establish a connection between the current small node and the node with the best score.

4. The novel recommendation method based on a neural network platform and node feature enhancement according to claim 3 is characterized in that: The specific process of step 4 is as follows: Step 4.1: Obtain the feature embedding matrix of different meta-paths after fusion back, After fusion Metapath The feature embedding matrix of , calculates the representation ability score of different meta-paths: (11)； in, It is Metapath The expression ability score of is the set of all nodes in the heterogeneous graph; is the attention vector; and are different learnable parameters; It is nodes In the Metapath The feature embedding vector of Step 4.2: Normalize the representation ability score to obtain the attention weight: (12)； in, It is Metapath The attention weight of is the number of meta-paths; It is Metapath The expression ability score of Step 4.

3. According to the attention weight, obtain the final embedding matrix representation of the node: (13)； in, It is nodes The final embedding matrix representation of ; It is Metapath The attention weight of .

5. The novel recommendation method based on a neural network platform and node feature enhancement according to claim 4 is characterized in that: In step 5, the entire model is optimized and trained by minimizing the following loss function: (14)； in, is the loss function value; is a set of nodes; A node collection middle Node of order; for The true label of is the classifier parameter; for The final embedding matrix representation of .

6. The novel recommendation method based on a neural network platform and node feature enhancement according to claim 5 is characterized in that: In step 6, for the user node, the final embedding matrix representation of the user node is obtained according to formula (13); for the novel node, the final embedding matrix representation of the novel node is obtained; Perform similarity calculation to 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 A personalized recommendation list is generated for the novels corresponding to the similarity values. At this time, the recommendation list contains the novels that best match the user's preferences.

Images