CN-122021720-A - Dynamic optimization method and system for graph neural network fusing multi-semantic hypergraph attention

CN122021720ACN 122021720 ACN122021720 ACN 122021720ACN-122021720-A

Abstract

The invention provides a graph neural network dynamic optimization method and system integrating multiple semantic hypergraph attention, relates to the technical field of intelligent information processing and complex system modeling, and aims at solving the problem that the prior art is difficult to realize multiple semantic high-order associated modeling and dynamic self-adaptive feature updating at the same time. The method comprises the steps of obtaining multidimensional data of entities in a target system, constructing a node characteristic matrix, a hypergraph structure containing multiple types of hyperedges and a hypergraph neural network characteristic propagation model, obtaining characteristic representation after node updating through a multi-head hypergraph attention mechanism, calculating a time feedback item by adopting a time-varying feedback mechanism through fusing a node historical state, local dynamic characteristics and a hypergraph propagation result, dynamically updating the node state, and generating a node high-dimensional embedded vector based on the characteristic representation after node updating and the node state after dynamic updating. The invention solves the problems existing in the prior art and improves the feature expression and dynamic optimization capability of the system in a complex structural scene.

Inventors

LIU HONG
LI WENHAO
FAN BAOYU
LI XIAOCHUAN
DU PING

Assignees

山东师范大学
浪潮电子信息产业股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (10)

1. The dynamic optimization method of the graph neural network integrating the multi-semantic hypergraph attention is characterized by comprising the following steps of: acquiring multidimensional data of an entity in a target system, and constructing a node characteristic matrix, wherein the multidimensional data at least comprises a space position, a behavior characteristic, a semantic attribute, environment information and a relationship characteristic; constructing a hypergraph structure containing multiple types of hyperedges according to the node characteristic matrix; Based on the hypergraph structure, constructing a hypergraph neural network characteristic propagation model; based on a hypergraph neural network feature propagation model, obtaining a feature representation after node updating through a multi-head hypergraph attention mechanism; calculating a time feedback item by adopting a time-varying feedback mechanism and fusing the historical state, the local dynamic characteristics and the hypergraph propagation result of the node; dynamically updating the node state based on the updated characteristic representation and the time feedback item; And generating a node high-dimensional embedded vector based on the node updated feature representation and the dynamically updated node state.
2. The dynamic optimization method for the graph neural network fusing the attention of the multi-semantic hypergraph according to claim 1, wherein the node feature matrix is constructed by the following specific processes: Performing time synchronization and normalization processing on the multidimensional data of the entity to obtain an original feature set; uniformly encoding the original feature set to obtain a node feature vector; performing feature standardization and dimension alignment treatment on the node feature vectors to form a node feature matrix; and carrying out unified normalization and time synchronization processing on various features of the node feature matrix to obtain the processed node feature vector representation.
3. The method of claim 1, wherein the multiple types of hyperedges include at least hyperedges generated based on spatial proximity, visual reachability, path sharing, mental similarity, and social relationships.
4. The method for dynamically optimizing a graph neural network by fusing multi-semantic hypergraph attention as recited in claim 1, wherein the hypergraph structure comprising multi-type hyperedges is constructed by the following specific processes: Setting a node set of a scene, and representing the processed node characteristic vector as a node; Based on the multidimensional feature similarity and interaction relation among the nodes, a superside set is constructed, based on the node set and the node feature vector representation, a space adjacent superside set, a view reachable superside set, a path sharing superside set, a psychological similar superside set and a social relation superside set are respectively generated, and the comprehensive high-order interaction set is obtained through combination; constructing a node-superside incidence matrix based on the node set and the superside set of the corresponding type; Calculating the superside weight according to the average similarity of the nodes in the superside to obtain a superside weight diagonal matrix; calculating a node degree matrix and a superside degree matrix based on the node-superside incidence matrix and the superside weight diagonal matrix; And integrating the node set, the comprehensive high-order interaction set, the node-superside incidence matrix, the superside weight diagonal matrix, the node degree matrix and the superside degree matrix to obtain a hypergraph structure of the multi-type superside.
5. The dynamic optimization method of the graph neural network fusing the multi-semantic hypergraph attention as claimed in claim 1, wherein the characteristic representation after node update is obtained through a multi-head hypergraph attention mechanism comprises the following specific processes: for each attention header, defining a node-over-edge attention weight; Based on the node-superside attention weight, calculating the attention weight of the node in the corresponding superside through the characteristic correlation between the node and the superside; Based on the attention weight of the node in the corresponding superside, carrying out weighted aggregation on the corresponding superside characteristics; and fusing sentence aggregation results of different attention heads to obtain the characteristic representation after node updating.
6. The method for dynamically optimizing a graph neural network by fusing multi-semantic hypergraph attention according to claim 1, wherein a time feedback term calculation formula is as follows: ; Wherein, the Is a change in speed; The local density change reflects the change of the congestion degree around the node; reflecting the risk potential energy change; is a set of 3 adjustment coefficients, which are set to be equal, And (2) and 。
7. The method for dynamically optimizing a graph neural network by fusing multi-semantic hypergraph attention according to claim 1, wherein the node characteristic dynamic update formula is: ; Wherein, the The node semantic features obtained by the concentration of the multi-head hypergraph reflect the current structural information; The characteristic mapping matrix is used for mapping the characteristic at the previous moment to the current updating space and is a learnable parameter; is a historical state retention factor for adjusting model memory capacity; the time feedback influence coefficient is used for controlling the action intensity of the dynamic feedback item on node updating; is a time feedback term used for reflecting the dynamic change trend of the node.
8. The graph neural network dynamic optimization system integrating the multi-semantic hypergraph attention is characterized by comprising the following components: The data acquisition and construction module is used for acquiring multidimensional data of the entity in the target system and constructing a node characteristic matrix, wherein the multidimensional data comprises a space position, a behavior characteristic, a semantic attribute, environmental information and a relationship characteristic; The hypergraph structure construction module is used for constructing a hypergraph structure containing multiple types of hyperedges according to the node characteristic matrix; the multi-semantic attention propagation module is used for constructing a hypergraph neural network characteristic propagation model based on the hypergraph structure; the method is used for obtaining the characteristic representation after node updating through a multi-head hypergraph attention mechanism based on the hypergraph neural network characteristic propagation model; the time-varying dynamic feedback updating module is used for calculating a time feedback item by adopting a time-varying feedback mechanism and fusing the historical state, the local dynamic characteristics and the supergraph propagation result of the node; dynamically updating the node state based on the updated characteristic representation and the time feedback item; And the node embedding generation module is used for generating a node high-dimensional embedding vector based on the node updated characteristic representation and the node state after dynamic update.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the method of any of the preceding claims 1-7 when the program is executed.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, performs the steps of the method of any of the preceding claims 1-7.

Description

Dynamic optimization method and system for graph neural network fusing multi-semantic hypergraph attention Technical Field The invention belongs to the technical field of intelligent information processing and complex system modeling, and particularly relates to a dynamic optimization method and a dynamic optimization system for a graph neural network by fusing multi-semantic hypergraph attention. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. Along with the continuous expansion of the scale of a complex system and the increasing complexity of an interaction mode among entities, the traditional graph structure model has obvious limitations in the aspects of representing high-order nonlinear association, fusing multi-source heterogeneous semantic information, adapting to dynamic evolution process and the like. In particular, in application scenarios involving multi-agent collaboration, high-dimensional space-time behavior modeling, and structured data depth mining, the entities generally have multi-dimensional correlation properties such as spatial adjacency, semantic similarity, relationship dependence, and state time variability, and if only a neural network model with a single graph structure or a fixed topology is used, it is difficult to comprehensively and accurately describe multi-dimensional, multi-level, multi-temporal structured features and evolution rules contained in a complex system In recent years, a graph neural network (Graph Neural Network, GNN) has made remarkable progress in capturing relationships between nodes and feature propagation, however, it is generally constructed based on a common graph structure, and only can represent binary relationships, and cannot describe high-order relationships between multiple entities, which is unfavorable for modeling of complex groups or multi-semantic information. In addition, the existing GNN model is fixed in updating mechanism in a dynamic scene, is difficult to adapt to nonlinear evolution of entity state change along with time, and lacks effective self-adaptive feedback capability. At present, a hypergraph structure is introduced into feature modeling, and expression of a high-order association relationship is realized by connecting a plurality of nodes through hyperedges. However, the conventional hypergraph method still has limitations in terms of dimensionality of semantic expression, distribution of attention weights, fusion of multi-source features, dynamic feedback update and the like, and is difficult to consider the requirement of high-order relational modeling under multi-semantic context. Meanwhile, the existing dynamic optimization method has insufficient response to the historical state, and cannot realize the self-adaptive adjustment of the information propagation path and the characteristic update. Disclosure of Invention In order to overcome the defects in the prior art, the invention provides a graph neural network dynamic optimization method and a system for fusing the attention of the multi-semantic hypergraph. The method comprises the steps of constructing a hypergraph structure containing multiple semantic relations, introducing a multi-head hypergraph attention mechanism, realizing weighted aggregation of multidimensional association features between nodes and hyperedges, and enabling a model to carry out self-adaptive adjustment according to historical states, risk changes and semantic contexts through a time-varying feedback and dynamic updating mechanism, so that feature expression capacity and dynamic optimization capacity of a system under a complex structural scene are improved. To achieve the above object, one or more embodiments of the present invention provide the following technical solutions: The first aspect of the invention provides a graph neural network dynamic optimization method fusing multi-semantic hypergraph attention, comprising the following steps: acquiring multidimensional data of an entity in a target system, and constructing a node characteristic matrix, wherein the multidimensional data comprises spatial positions, behavior characteristics, semantic attributes, environmental information and relationship characteristics; constructing a hypergraph structure containing multiple types of hyperedges according to the node characteristic matrix; Based on the hypergraph structure, constructing a hypergraph neural network characteristic propagation model; based on a hypergraph neural network feature propagation model, obtaining a feature representation after node updating through a multi-head hypergraph attention mechanism; calculating a time feedback item by adopting a time-varying feedback mechanism and fusing the historical state, the local dynamic characteristics and the hypergraph propagation result of the node; dynamically updating the node state based on the updated characteristic represe