A graph neural network-based graph data processing method

Abstract

This invention provides a graph data processing method based on graph neural networks, aiming to solve the problems of excessive smoothness, difficulty in capturing long-range dependencies, and sensitivity to noise in the original graph structure of existing graph neural networks. Its core includes: multimodal feature extraction and fusion of the original graph data to form enhanced feature representations of nodes; constructing an innovative graph neural network model, which sequentially includes a bidirectional information propagation layer, a hierarchical attention fusion layer, and an adaptive graph structure learning layer; using this model for information processing, where the bidirectional propagation layer simultaneously performs message passing from node to neighbor and from neighbor to node, the attention fusion layer dynamically integrates features at different depths of the network, and the structure learning layer adaptively optimizes the graph topology; finally, the model is trained through a multi-task joint learning framework to complete downstream tasks. This invention significantly improves the accuracy and robustness of graph data processing tasks.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer information technology, and in particular to an improved graph data processing method based on graph neural networks (GNN). Background Technology

[0002] Graph-structured data is widely present in the real world, such as in social networks, recommender systems, molecular chemistry, and knowledge graphs. Graph neural networks are powerful tools for processing this type of data; they learn node and graph representations through message passing between neighboring nodes.

[0003] However, existing typical graph neural network models have several inherent defects: First, the problem of oversmoothing: as the number of network layers increases, the representations of all nodes tend to be the same, making it impossible for the model to distinguish different nodes; Second, difficulty in capturing long-range dependencies: nodes have difficulty effectively acquiring information about key neighbors far away from their multi-hops; Third, sensitivity to noise in the original graph structure: the model performance is too dependent on the quality of the graph structure of the initial input, and the performance will drop significantly if there are missing or incorrect connections in the graph structure.

[0004] Therefore, there is an urgent need for a new graph data processing method that can learn the deep representations of graph data more effectively and robustly. Summary of the Invention

[0005] (a) Technical problems to be solved In view of the above-mentioned technical problems, the present invention provides a graph data processing method based on graph neural networks.

[0006] (II) Technical Solution This invention provides a graph data processing method based on graph neural networks, comprising the following steps: Step 1: Obtain the original graph data, and perform preprocessing and multimodal feature fusion on the original graph data to obtain the enhanced initial feature representation of each node; Step 2: Construct a graph neural network model, which sequentially includes a bidirectional information propagation layer, a hierarchical attention fusion layer, and an adaptive graph structure learning layer; Step 3: Input the enhanced initial feature representation into the graph neural network model for processing; The bidirectional information propagation layer is used to perform message passing calculations in two directions. The first direction is for a node to pass its own state to its neighboring nodes, and the second direction is for neighboring nodes to converge information to the central node, thereby generating a preliminary embedded representation of the node containing bidirectional context. The hierarchical attention fusion layer is used to receive and integrate the node output representations of each layer in the bidirectional information propagation layer. It assigns different importance weights to the representations of each layer through an attention mechanism, and then weights and fuses the representations of all layers to form the final fused representation of each node. The adaptive graph structure learning layer is used to dynamically calculate the new association strength between nodes based on the final fused representation, and generate a task-driven, optimized new adjacency matrix. Step 4: Based on the objectives of the downstream tasks, calculate the loss using the output of the graph neural network model, and train and optimize the model parameters using a multi-task joint learning framework.

[0007] In some embodiments of the present invention, the preprocessing and multimodal feature fusion in step one includes: encoding and standardizing the original features of the nodes themselves; extracting the type and weight features of the edges; generating global features that describe the statistical characteristics of the entire graph; and finally fusing the node features, edge features and global features into a unified node feature vector by splicing or an attention-based mechanism.

[0008] In some embodiments of the present invention, the message function and aggregation function in the bidirectional information propagation layer may employ, but are not limited to, mean pooling, max pooling, attention-based weighted sums, or gating mechanisms.

[0009] In some embodiments of the present invention, the attention weights in the hierarchical attention fusion layer are jointly determined by the layer representation of each node and a global context vector, and are dynamically calculated through a learnable parameter matrix and a nonlinear activation function, ensuring that the fusion process depends on the specific node and task context.

[0010] In some embodiments of the present invention, the adaptive graph structure learning layer calculates the cosine similarity or dot product of the feature representations of each pair of nodes after nonlinear transformation, then generates new connection weights between nodes through an activation function, and combines the learned new adjacency matrix with the original adjacency matrix in a learnable manner.

[0011] In some embodiments of the present invention, the total loss function of the multi-task joint learning framework described in step four is composed of a weighted sum of three parts: the main task loss, the self-supervised loss based on graph structure reconstruction, and the sparsity constraint loss on the newly generated adjacency matrix. (III) Beneficial Effects As can be seen from the above technical solution, the present invention has the following beneficial effects: This invention fuses shallow and deep features through a hierarchical attention mechanism, avoiding the oversmoothing problem caused by using only the deepest features. The bidirectional propagation mechanism and deep network structure facilitate the flow and interaction of information over longer distances within the graph. The adaptive graph structure learning layer dynamically learns and optimizes the topology, reducing the model's sensitivity to noise and missing connections in the original input graph, thus improving its robustness and generalization ability. Through the synergistic work of these innovative modules, the method of this invention is expected to achieve higher accuracy and stability on various graph learning tasks. Attached Figure Description

[0012] Figure 1 This is an overall flowchart of the graph data processing method of the present invention. Detailed Implementation

[0013] This invention provides a graph data processing method based on graph neural networks. To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0014] Specific embodiments: such as Figure 1 As shown, this invention provides a graph data processing method based on graph neural networks, the steps of which are as follows: Step 1: Obtain the original graph data, and perform preprocessing and multimodal feature fusion on the original graph data to obtain the enhanced initial feature representation of each node; Step 2: Construct a graph neural network model, which sequentially includes a bidirectional information propagation layer, a hierarchical attention fusion layer, and an adaptive graph structure learning layer; Step 3: Input the enhanced initial feature representation into the graph neural network model for processing; The bidirectional information propagation layer is used to perform message passing calculations in two directions. The first direction is for a node to pass its own state to its neighboring nodes, and the second direction is for neighboring nodes to converge information to the central node, thereby generating a preliminary embedded representation of the node containing bidirectional context. The hierarchical attention fusion layer is used to receive and integrate the node output representations of each layer in the bidirectional information propagation layer. It assigns different importance weights to the representations of each layer through an attention mechanism, and then weights and fuses the representations of all layers to form the final fused representation of each node. The adaptive graph structure learning layer is used to dynamically calculate the new association strength between nodes based on the final fused representation, and generate a task-driven, optimized new adjacency matrix. Step 4: Based on the objectives of the downstream tasks, calculate the loss using the output of the graph neural network model, and train and optimize the model parameters using a multi-task joint learning framework.

[0015] The preprocessing and multimodal feature fusion described in step one includes encoding and standardizing the original features of the nodes themselves; extracting the type and weight features of the edges; generating global features that describe the statistical characteristics of the entire graph; and finally, fusing the node features, edge features, and global features into a unified node feature vector through concatenation or attention-based mechanisms.

[0016] The message functions and aggregation functions in the bidirectional information propagation layer may employ, but are not limited to, mean pooling, max pooling, attention-based weighted sums, or gating mechanisms.

[0017] The attention weights in the hierarchical attention fusion layer are determined by the layer representation of each node and a global context vector. They are dynamically calculated using a learnable parameter matrix and a nonlinear activation function, ensuring that the fusion process depends on the specific node and task context.

[0018] The adaptive graph structure learning layer calculates the cosine similarity or dot product of the feature representations of each pair of nodes after nonlinear transformation, then uses an activation function to generate new connection weights between nodes, and combines the learned new adjacency matrix with the original adjacency matrix in a learnable manner.

[0019] The multi-task joint learning framework described in step four has a total loss function composed of three parts in a weighted sum: the main task loss, the self-supervised loss based on graph structure reconstruction, and the sparsity constraint loss on the newly generated adjacency matrix.

[0020] This concludes the detailed description of the embodiment with reference to the accompanying drawings. Based on the above description, those skilled in the art should have a clear understanding of the present invention.

[0021] It should be noted that implementations not shown or described in the accompanying drawings or the main text of the specification are all forms known to those skilled in the art and are not described in detail. Furthermore, the definitions of the elements and methods described above are not limited to the various specific structures, shapes, or methods mentioned in the embodiments.

[0022] It should also be noted that this document provides examples of parameters containing specific values, but these parameters need not be exactly equal to the corresponding values, but can approximate the corresponding values ​​within acceptable error tolerances or design constraints. Furthermore, unless specifically described or steps must occur in sequence, the order of the above steps is not limited to those listed above and can be varied or rearranged according to the desired design. Moreover, the above embodiments can be used in combination with each other or with other embodiments based on design and reliability considerations; that is, technical features from different embodiments can be freely combined to form more embodiments.

[0023] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A graph data processing method based on graph neural networks, characterized in that, Includes the following steps: Step 1: Obtain the original graph data, and perform preprocessing and multimodal feature fusion on the original graph data to obtain the enhanced initial feature representation of each node; Step 2: Construct a graph neural network model, which sequentially includes a bidirectional information propagation layer, a hierarchical attention fusion layer, and an adaptive graph structure learning layer; Step 3: Input the enhanced initial feature representation into the graph neural network model for processing; The bidirectional information propagation layer is used to perform message passing calculations in two directions. The first direction is for a node to pass its own state to its neighboring nodes, and the second direction is for neighboring nodes to converge information to the central node, thereby generating a preliminary embedded representation of the node containing bidirectional context. The hierarchical attention fusion layer is used to receive and integrate the node output representations of each layer in the bidirectional information propagation layer. It assigns different importance weights to the representations of each layer through an attention mechanism, and then weights and fuses the representations of all layers to form the final fused representation of each node. The adaptive graph structure learning layer is used to dynamically calculate the new association strength between nodes based on the final fused representation, and generate a task-driven, optimized new adjacency matrix. Step 4: Based on the objectives of the downstream tasks, calculate the loss using the output of the graph neural network model, and train and optimize the model parameters using a multi-task joint learning framework.

2. The method according to claim 1, characterized in that, The preprocessing and multimodal feature fusion described in step one includes: encoding and standardizing the original features of the nodes themselves; extracting the type and weight features of the edges; generating global features that describe the statistical characteristics of the entire graph; and finally fusing the node features, edge features, and global features into a unified node feature vector through concatenation or attention-based mechanisms.

3. The method according to claim 1, characterized in that, The message functions and aggregation functions in the bidirectional information propagation layer may employ, but are not limited to, mean pooling, max pooling, attention-based weighted sums, or gating mechanisms.

4. The method according to claim 1, characterized in that, The attention weights in the hierarchical attention fusion layer are determined by the layer representation of each node and a global context vector. They are dynamically calculated using a learnable parameter matrix and a nonlinear activation function, ensuring that the fusion process depends on the specific node and task context.

5. The method according to claim 1, characterized in that, The adaptive graph structure learning layer calculates the cosine similarity or dot product of the feature representations of each pair of nodes after nonlinear transformation, then uses an activation function to generate new connection weights between nodes, and combines the learned new adjacency matrix with the original adjacency matrix in a learnable manner.

6. The method according to claim 1, characterized in that, The multi-task joint learning framework described in step four has a total loss function composed of three parts in a weighted sum: the main task loss, the self-supervised loss based on graph structure reconstruction, and the sparsity constraint loss on the newly generated adjacency matrix.

Images