Node representation learning method fusing diffusion hypergraph modeling and graph convolutional network

Abstract

The invention relates to a node representation learning method fusing diffusion hypergraph modeling and a graph convolutional network, and belongs to the technical field of modeling and prediction of a complex diffusion process. The method comprises the following steps: extracting a static node structure representation of a social graph, and obtaining a social context feature representation of the static node structure representation; constructing a diffusion hypergraph according to the propagation diffusion sequence, and dividing the diffusion hypergraph into a plurality of sub diffusion hypergraphs; performing feature extraction and fusion on each sub-diffusion hypergraph to obtain user information features; obtaining time sequence feature information embedding based on the improved LSTM network and the user information features; and performing cross-dimension splicing on the social context feature representation and the time sequence feature information to obtain a splicing fusion feature, obtaining a propagation probability based on the splicing fusion feature, and completing model performance evaluation based on the propagation probability. The objective of the invention is to solve the technical problems that time sequence information is not deeply learned and global information is not integrated in the prior art.

Description

Technical Field

[0001] This invention relates to a node representation learning method that integrates diffusion hypergraph modeling and graph convolutional networks, belonging to the field of modeling and prediction technology for complex diffusion processes. Background Technology

[0002] In fields such as social networks and news dissemination, information diffusion modeling has received widespread attention as an important technique for understanding user behavior, predicting information propagation paths, and evaluating intervention strategies. Traditional information diffusion modeling methods mainly rely on rule-driven approaches such as propagation probability modeling and infection threshold models (e.g., IC and LT models), but in practical applications, they often struggle to capture complex temporal dependencies and high-order structural features.

[0003] In recent years, with the development of Graph Neural Network (GNN) technology, researchers have begun to apply it to diffusion modeling tasks. Graph Neural Networks excel at processing graph-structured data and can effectively extract local and global semantic information of nodes. In particular, Graph Convolutional Networks (GCNs) achieve deep learning of node representations by aggregating information from neighboring nodes.

[0004] However, simply using static graph modeling methods such as GCN is insufficient to fully capture the temporal information and dynamic evolutionary characteristics of diffusion sequences. To address this, researchers have introduced sequence modeling techniques, such as Long Short-Term Memory (LSTM) networks, to capture the temporal order and dependencies between nodes along the diffusion path. Furthermore, to simulate the real process of information propagating from one node to multiple nodes along a path, a higher-order data structure—Diffusion Hypergraph—has gradually become a research hotspot. A diffusion hypergraph connects multiple nodes with hyperedges, enabling a more natural expression of diverse propagation relationships and multi-hop diffusion structures. By mapping diffusion sequences to a hypergraph structure and combining GCN and LSTM models, the structural and temporal characteristics involved in the diffusion process can be modeled more effectively.

[0005] While these deep learning-based models offer significant advantages in uncovering latent features in information and reducing human intervention and time costs, two key issues remain to be addressed. First, they fail to deeply learn temporal features; most models, such as Recurrent Neural Networks (RNNs) and their variants, simply extract temporal information from the cascading process without deeply learning temporal sequence information. Second, user information is extracted using Graph Convolutional Networks (GCNs), aggregating information from neighboring nodes, thus only extracting local information and never integrating global information. Summary of the Invention

[0006] The purpose of this invention is to provide a node representation learning method that integrates diffuse hypergraph modeling and graph convolutional networks, aiming to solve the technical problems of existing technologies that do not deeply learn temporal information and do not integrate global information.

[0007] To achieve the above objectives, the technical solution of this invention is: a node representation learning method that integrates diffuse hypergraph modeling and graph convolutional networks, the specific steps of which are as follows:

[0008] Step 1: Extract the static node structure representation of the social graph and obtain the social context feature representation of the static node structure representation;

[0009] Step 2: Construct a diffusion hypergraph based on the propagation and diffusion sequence, and divide the diffusion hypergraph into several sub-diffusion hypergraphs;

[0010] Step 3: Extract and fuse features from each sub-diffusion hypergraph to obtain user information features;

[0011] Step 4: Obtain temporal feature information embedding based on the improved LSTM network and the user information features; wherein, the improved LSTM network is an attention mechanism introduced into the original LSTM network;

[0012] Step 5: Perform cross-dimensional concatenation of the social context feature representation and the temporal feature information to obtain the concatenated and fused feature. Based on the concatenated and fused feature, obtain the propagation probability and complete the model performance evaluation based on the propagation probability.

[0013] Optionally, Step 1 includes:

[0014] Step 1.1: Define the graph composed of social networks as a social graph. ,in Let n represent the set of users. The number of users in China Let m be the set of edges, representing the connections between users in a social network. The number of sides in the middle;

[0015] Step 1.2: Extract the static node structure representation of the social graph using a Graph Convolutional Network (GCN). The expression is:

[0016]

[0017] in, This is a static node structure representation. It is the ReLU function. Trainable parameter matrix, and These are social graphs The adjacency matrix and degree matrix, The embedding is initialized using a random normal distribution, where d is the dimension of the embedding. yes The number of users in China;

[0018] Step 1.3: Represent the static node structure Represented as ,in, for One of the user characteristics is represented by the time decay effect, for the last m users propagated. The feature representations are then subjected to mean pooling to obtain the social context feature representation, expressed as follows:

[0019]

[0020] in, For social context features, mean() means taking the mean.

[0021] Optionally, Step 2 includes:

[0022] Step 2.1: Construct a set of hyperedges based on the propagation and diffusion sequence In this context, a hyperedge represents a message and consists of multiple groups of users and their timestamps, expressed as follows: , It is a timestamp, and satisfies < , Indicates the maximum number of times a message can be forwarded;

[0023] Step 2.2: Form a diffuse hypergraph from all hyperedges. The diffusion hypergraph is divided into T sub-diffusion hypergraphs according to the timestamp and a preset step size, as expressed by:

[0024]

[0025]

[0026] in, For sub-diffusion hypergraph, It is a collection of users. It is a set of superedges.

[0027] Optionally, Step 3 includes:

[0028] Step 3.1: Define a hyperedge in the hyperedge set as... Aggregate all with hyperedge Information from adjacent nodes yields the features of the hyperedge. The expression is:

[0029]

[0030] in, It is the ReLU function. It is the aggregation of nodes with hyperedge adjacency. For nodes, For neighbor node i at time t, the first... The input features of the layer These are weight values;

[0031] Step 3.2: Update nodes Features The expression is:

[0032]

[0033] in, It is the ReLU function. For the updated features, It is a node All connected hyperedges, It is the first The trainable parameters of the layer;

[0034] Step 3.3: Introduce the Hypergraph Attention Network (HGAN) for feature fusion, the expression is:

[0035]

[0036] in, It is a node initial characteristics, From the child diffusion hypergraph Nodes in the first layer of HGAN The output embedding, It is a transformation matrix. It is a vector used to calculate the attention score, thereby obtaining the final user information features. .

[0037] Optionally, Step 4 includes:

[0038] Step 4.1: Analyze user information features The input is fed into the improved LSTM network to obtain the hidden state at the current time step, expressed as:

[0039]

[0040] in, It is in a hidden state. User information features at the current time step , These are the parameters in LSTM, which include the input gate, forget gate, memory unit, output gate, and candidate state;

[0041] Step 4.2: Hiding the state for all time steps Weighted summarization is performed to obtain the embedded time-series feature information. The expression is:

[0042]

[0043] in, This represents the attention weight.

[0044] Optionally, the attention weights are dynamically calculated based on an attention score function, the expression of which is:

[0045]

[0046] in, To score attention, Here is the weight matrix, used to project the hidden state into the attention space; For bias terms, The hyperbolic tangent activation function is used. Score mapping vector;

[0047] Obtain attention weights using softmax .

[0048] Optionally, the propagation probability expression is:

[0049]

[0050] in, For the probability of propagation, For activation function, It's a concatenation operation. It is a weight matrix. It is a deviation variable.

[0051] The beneficial effects of this invention are:

[0052] 1. This invention introduces a structure modeling approach based on diffusion hypergraphs, combining graph convolutional networks (GCN) and LSTM networks to achieve joint modeling of temporal dependencies and topological structures between nodes in information diffusion paths, significantly improving the ability to express complex information propagation behavior.

[0053] 2. Mean Pooling is used to fuse local graph structure information and combined with a time series model to extract diffusion sequences, thereby obtaining a more refined and dynamic node embedding representation, which effectively improves the accuracy of subsequent tasks such as node classification, link prediction or diffusion behavior prediction.

[0054] 3. The modeling framework proposed in this invention can adapt to data graphs of different sizes and structures, and can be flexibly extended to other graph neural network models or temporal modeling units (such as Transformer), which has strong versatility and engineering application value. Attached Figure Description

[0055] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1: As Figure 1 As shown, a node representation learning method that integrates diffuse hypergraph modeling and graph convolutional networks includes the following steps:

[0058] Step 1: Extract the static node structure representation of the social graph and obtain the social context feature representation of the static node structure representation;

[0059] It's important to understand that in real-world social interactions, individuals tend to establish social connections with people who share similar preferences, leading to forwarding behavior. This phenomenon is known as social homophily. For example, friends with similar interests, opinions, or behavioral habits are more likely to influence each other. Therefore, user social homophily can be reflected in the structure of social networks. Thus, to better model the static dependencies between users, this embodiment introduces a social graph to represent social connections and utilizes a multi-layer graph convolutional network (GCN) to model it, thereby learning user embeddings with social semantics.

[0060] Step 1.1: Define the graph composed of social networks as a social graph. ,in Let n represent the set of users. The number of users in China Let m be the set of edges, representing the connections between users in a social network. The number of sides in the middle;

[0061] Step 1.2: Extract the static node structure representation of the social graph using a Graph Convolutional Network (GCN). The expression is:

[0062]

[0063] in, This is a static node structure representation. It is the ReLU function. Trainable parameter matrix, and These are social graphs The adjacency matrix and degree matrix, The embedding is initialized using a random normal distribution, where d is the dimension of the embedding. yes The number of users in China; This represents the output of the first layer of the graph convolutional network. This indicates the graph convolution output of the second layer. This invention employs a two-layer graph convolutional network, therefore... .

[0064] Step 1.3: Represent the static node structure Represented as ,in, for One of the user characteristics is represented by the time decay effect, for the last m users propagated. The feature representations are then subjected to mean pooling to obtain the social context feature representation, expressed as follows:

[0065]

[0066] in, For social context features, mean() means taking the mean.

[0067] It is important to understand that, in order to reduce the bias in user feature representation, this embodiment performs mean normalization on the user feature representation. During information dissemination, the next disseminator of information is often influenced by the combined actions of users who recently forwarded the information; that is, its dissemination decision depends not only on a single upstream user, but is driven by the combined behavior of a group of previous disseminators. Therefore, for user feature representation... Since users most recently propagated are generally more affected in micro-cascade diffusion propagation prediction (i.e., due to the time decay effect), this embodiment analyzes the last m users propagated. The feature representations are averaged and pooled to obtain the social context feature representations.

[0068] Specifically, let's further illustrate this with a concrete example. Assuming each user has a feature dimension of 2 and follows m=3 users (time decay window), the calculation process for the mean pooling result is as follows:

[0069] The last 3 users: { , , }={ , , };

[0070] The corresponding features are: =[0.8,0.2], =[0.5,0.7], =[0.1,0.9];

[0071] The mean pooling result is as follows: =[ , = [0.467, 0.6].

[0072] Step 2: Construct a diffusion hypergraph based on the propagation and diffusion sequence, and divide the diffusion hypergraph into several sub-diffusion hypergraphs;

[0073] Step 2.1: Construct a set of hyperedges based on the propagation and diffusion sequence In this context, a hyperedge represents a message and consists of multiple groups of users and their timestamps, expressed as follows: , It is a timestamp, and satisfies < , Indicates the maximum number of times a message can be forwarded;

[0074] Step 2.2: Form a diffuse hypergraph from all hyperedges. The diffusion hypergraph is divided into T sub-diffusion hypergraphs according to the timestamp and a preset step size, as expressed by:

[0075]

[0076]

[0077] in, For sub-diffusion hypergraph, It is a collection of users. It is a set of superedges.

[0078] Understandably, to simultaneously model the global interaction relationships between cascades and their dynamic characteristics over time, this embodiment uses a constructed diffusion hypergraph to capture the global dependencies of users at the cascade level within each time interval. Based on this, a cross-time fusion layer is designed to model the structural changes between adjacent time intervals, thereby characterizing the temporal evolution dynamics of the diffusion process.

[0079] Step 3: Extract and fuse features from each sub-diffusion hypergraph to obtain user information features;

[0080] Step 3.1: Define a hyperedge in the hyperedge set as... Aggregate all with hyperedge Information from adjacent nodes yields the features of the hyperedge. The expression is:

[0081]

[0082] in, It is the ReLU function. It is the aggregation of nodes with hyperedge adjacency. For nodes, For neighbor node i at time t, the first... The input features of the layer As the weight value, in this embodiment, ;

[0083] Step 3.2: Update nodes Features The expression is:

[0084]

[0085] in, For the updated features, It is a node All connected hyperedges, It is the first The trainable parameters of the layer;

[0086] Step 3.3: Since Steps 3.1 and 3.2 only learn information from their respective time periods and cannot represent the information of the diffusion sequence, a Hypergraph Attention Network (HGAN) is introduced for feature fusion. The expression is:

[0087]

[0088] in, It is a node initial characteristics, From the child diffusion hypergraph Nodes in the Lth layer of HGAN The output embedding, It is a transformation matrix. It is a vector used to calculate the attention score, thereby obtaining the final user information features. .

[0089] Step 4: Obtain temporal feature information embedding based on the improved LSTM network and the user information features; wherein, the improved LSTM network is an attention mechanism introduced into the original LSTM network;

[0090] It is important to understand that, in order to further enhance the model's ability to model sequence characteristics, this embodiment improves the original LSTM network by introducing an attention mechanism to deeply extract time-sensitive behavioral patterns of users in the propagation path. Specifically, the attention mechanism can dynamically identify and strengthen the focus on key propagation nodes in the propagation path, thereby enhancing the LSTM's ability to model long-term dependencies and important user states.

[0091] Step 4.1: Analyze user information features The input is fed into the improved LSTM network to obtain the hidden state at the current time step, expressed as:

[0092]

[0093] in, It is in a hidden state. User information features at the current time step , These are the parameters in LSTM, including the input gate, forget gate, memory unit, output gate, and candidate state. The parameters are defined as follows:

[0094]

[0095] in, It is the sigmoid function. , These are trainable parameters. It refers to the user information characteristics from the previous moment. It is a deviation variable. It is the Gate of Oblivion. It is a memory unit. It is the output gate, and tanh is the hyperbolic tangent function. It is a candidate state. Candidate states are added to the cell state according to the input gate ratio to obtain the hidden state. .

[0096] Step 4.2: Hiding the state for all time steps Weighted summarization is performed to obtain the embedded time-series feature information. The expression is:

[0097]

[0098] in, This represents the attention weight.

[0099] The attention weights are dynamically calculated based on the attention score function, the expression of which is:

[0100]

[0101] in, To score attention, Here is the weight matrix, used to project the hidden state into the attention space; For bias terms, The activation function is hyperbolic tangent, ensuring gradient stability. The score mapping vector is used to compress the attention space output into a scalar score. ;

[0102] Obtain attention weights using softmax .

[0103] It is understandable that, in this embodiment, after the k-th time step, the hidden state output by the LSTM with attention mechanism represents the embedding of sequential feature information. It contains all historical information from the beginning of the sequence to the present.

[0104] Step 5: Perform cross-dimensional concatenation of the social context feature representation and the temporal feature information to obtain the concatenated and fused feature. Based on the concatenated and fused feature, obtain the propagation probability and complete the model performance evaluation based on the propagation probability.

[0105] Optionally, the propagation probability expression is:

[0106]

[0107] in, For the probability of propagation, For activation function, It's a concatenation operation. It is a weight matrix. It is a deviation variable.

[0108] Optionally, model performance evaluation based on the propagation probability includes training the model and evaluating model performance, wherein the objective of training the model is to maximize the log-likelihood of all training cascade sequences, defined as:

[0109]

[0110] in, Let be the objective function. It is all cascade numbers, express The first in dimension, It is the first Cascade sequence 1 node These are all the parameters within the model. Indicates the maximum number of times a message can be forwarded;

[0111] The model performance is evaluated using two metrics: average accuracy and hit rate. Average accuracy measures the ranking quality of the top k predictions and is defined as follows:

[0112]

[0113] Where map@k represents the average precision. Indicates the length of the cascade. This refers to the probability of correctly predicting the j-th position among the first k predictions. It is based on propagation probability. If the prediction is correct, then = 1, otherwise = 0.

[0114] The hit rate is defined as the ratio of actually activated nodes among the top k predicted nodes:

[0115]

[0116] Where hits@k is the hit rate;

[0117] Based on the specific implementation details, the effectiveness of the technical solution of the present invention will be demonstrated through experiments.

[0118] In this embodiment, the model is evaluated using the average precision (map@k) and hit rate (hits@k) of the first k results, where k takes the value [10, 50, 100].

[0119] Specifically, this experiment was conducted on a server (NVIDIA A100-SXM4-80GB). For each dataset, a random training, test, and validation set ratio of 8:1:1 was used, with a maximum concatenation length of 200. PyTorch was used for validation and testing, with the Adam optimizer and a learning rate of 0.001. The dropout ratio was set to 0.3, the batch size to 64, the initial embedding dimension and hidden state dimension to 64, and the period to 50. In the experiment, a two-layer GCN was used for user feature extraction from the social graph; a single-layer HGAN was used for feature extraction during information diffusion and propagation.

[0120] Specifically, this experiment was conducted on two datasets: Android and Douban. Both datasets are derived from data available on social networking platforms. Android data comes from the Stack Exchanges Q&A community, where users establish friendships through interactive behaviors such as asking and answering questions. Douban data comes from a social networking site where users share their reading status for books or movies; users who read the same books together are considered friends. Statistical information for the datasets is shown in Table 1. The experiment compares the performance of this invention with existing representative models, and the experimental results are shown in Tables 2 and 3.

[0121] Table 1. Statistical information of the dataset

[0122]

[0123] Table 2 Experimental results based on the Douban dataset (%)

[0124]

[0125] Table 3. Experimental results based on the Android dataset (%)

[0126]

[0127] Among them, the TopoLSTM model models the cascading diffusion process of information as a dynamic directed acyclic graph (DAG), extending the traditional LSTM structure to learn topology-aware user embeddings, thereby improving the predictive ability of message propagation. The NDM model constructs a micro-prediction model based on self-attention and convolutional neural networks to alleviate long-term dependency problems in prediction. The SNIDSA model is based on RNN and introduces a structural attention mechanism, improving the modeling effect of information propagation by combining the structural features of the diffusion path with cascading temporal information. The Inf-VAE model combines GNN to model the homogeneity features in social networks and achieves deep integration of social and temporal information through a collaborative attention mechanism. The TAN-DURD model simultaneously models the bidirectional dependency relationship between information senders and receivers to characterize the cascading propagation process of information.

[0128] As shown in the table, on the Douban dataset, this invention significantly outperforms existing methods across all metrics, particularly improving upon the best-performing comparison method, DyHGCN, by 20.01 percentage points on Hits@100, fully demonstrating the effectiveness of this invention. On the Android dataset, this invention also achieves optimal performance, especially on the MAP metric. This indicates that the model can not only more accurately predict users participating in the diffusion but also better identify core users in terms of ranking relevance.

[0129] In summary, this embodiment first constructs a time-aware diffusion hypergraph based on historical diffusion sequences to characterize one-to-many and many-to-many propagation relationships between users in different time intervals. Then, a hypergraph attention network is introduced to model the diffusion structure in different time slices, obtaining high-order semantic representations of users at each time step. To further capture the temporal dependencies in the diffusion path, an LSTM module with an attention mechanism is further introduced to achieve dynamic focusing of key nodes and time steps based on global structure modeling. Simultaneously, a GCN is used to extract static structural representations between users from the social graph, and mean pooling is used to compress global and local feature representations. Finally, multi-source feature fusion and a softmax classifier are used to predict the probability of the next propagation node. Experimental results on multiple real-world social datasets show that the proposed method has good prediction results, demonstrating good accuracy and generalization ability, validating the effectiveness of this invention in high-order structure and temporal sequence modeling.

[0130] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A node representation learning method that integrates diffuse hypergraph modeling and graph convolutional networks, characterized in that, The method includes the following steps: Step 1: Extract the static node structure representation of the social graph and obtain the social context feature representation of the static node structure representation; Step 2: Construct a diffusion hypergraph based on the propagation and diffusion sequence, and divide the diffusion hypergraph into several sub-diffusion hypergraphs; Step 3: Extract and fuse features from each sub-diffusion hypergraph to obtain user information features; Step 4: Obtain temporal feature information embedding based on the improved LSTM network and the user information features; wherein, the improved LSTM network is an attention mechanism introduced into the original LSTM network; Step 5: Perform cross-dimensional concatenation of the social context feature representation and the temporal feature information to obtain the concatenated and fused feature. Based on the concatenated and fused feature, obtain the propagation probability and complete the model performance evaluation based on the propagation probability.

2. The node representation learning method for fusing diffuse hypergraph modeling and graph convolutional networks according to claim 1, characterized in that, Step 1 includes: Step 1.1: Define the graph composed of social networks as a social graph. ,in Let n represent the set of users. The number of users in China Let m be the set of edges, representing the connections between users in a social network. The number of edges in the middle; Step 1.2: Extract the static node structure representation of the social graph using a Graph Convolutional Network (GCN). The expression is: ； ;in, This is a static node structure representation. It is the ReLU function. Trainable parameter matrix, and These are social graphs The adjacency matrix and degree matrix, The embedding is initialized using a random normal distribution, where d is the dimension of the embedding. yes The number of users in China; Step 1.3: Represent the static node structure Represented as ,in, for One of the user characteristics is represented by the time decay effect, for the last m users propagated. The feature representations are then subjected to mean pooling to obtain the social context feature representation, expressed as follows: ； in, For social context features, mean() means taking the mean.

3. The node representation learning method for fusing diffuse hypergraph modeling and graph convolutional networks according to claim 1, characterized in that, Step 2 includes: Step 2.1: Construct a set of hyperedges based on the propagation and diffusion sequence In this context, a hyperedge represents a message and consists of multiple groups of users and their timestamps, expressed as follows: , It is a timestamp, and satisfies < , Indicates the maximum number of times a message can be forwarded; Step 2.2: Form a diffuse hypergraph from all hyperedges. The diffusion hypergraph is divided into T sub-diffusion hypergraphs according to the timestamp and a preset step size, as expressed by: ； ； in, For sub-diffusion hypergraph, It is a collection of users. It is a set of superedges.

4. The node representation learning method for fusing diffuse hypergraph modeling and graph convolutional networks according to claim 1, characterized in that, Step 3 includes: Step 3.1: Define a hyperedge in the hyperedge set as... Aggregate all with hyperedge Information from adjacent nodes yields the features of the hyperedge. The expression is: ； in, It is the ReLU function. It is the aggregation of nodes with hyperedge adjacency. For nodes, For neighbor node i at time t, the first... The input features of the layer These are weight values; Step 3.2: Update nodes Features The expression is: ； in, For the updated features, It is a node All connected hyperedges, It is the first The trainable parameters of the layer; Step 3.3: Introduce the Hypergraph Attention Network (HGAN) for feature fusion, the expression is: ; ； in, It is a node initial characteristics, From the child diffusion hypergraph Nodes in the first layer of HGAN The output embedding, It is a transformation matrix. It is a vector used to calculate the attention score, thereby obtaining the final user information features. .

5. The node representation learning method for fusing diffuse hypergraph modeling and graph convolutional networks according to claim 1, characterized in that, Step 4 includes: Step 4.1: Analyze user information features The input is fed into the improved LSTM network to obtain the hidden state at the current time step, expressed as: ； in, It is in a hidden state. User information features at the current time step , These are the parameters in LSTM, which include the input gate, forget gate, memory unit, output gate, and candidate state; Step 4.2: Hidden states for all time steps Weighted summarization is performed to obtain the embedded time-series feature information. The expression is: ； in, For attention weights.

6. The node representation learning method for fusing diffuse hypergraph modeling and graph convolutional networks according to claim 5, characterized in that, The attention weights are dynamically calculated based on the attention score function, the expression of which is: ； in, To score attention, Here is the weight matrix, used to project the hidden state into the attention space; For bias terms, The hyperbolic tangent activation function is used. Score mapping vector; Obtain attention weights using softmax .

7. The node representation learning method for fusing diffuse hypergraph modeling and graph convolutional networks according to claim 1, characterized in that, The propagation probability expression is: ； in, For the probability of propagation, For activation function, It's a concatenation operation. It is a weight matrix. It is a deviation variable.

Images