CN-121984742-A - Hypergraph propagation source positioning method based on interaction enhancement

Abstract

The invention discloses a hypergraph propagation source positioning method based on interaction enhancement. The traditional tracing method mainly relies on paired interaction assumption of a simple graph structure, and the positioning accuracy under complex topology is limited. The method comprises the steps of firstly constructing a supergraph propagation model, obtaining a complete observation snapshot of network nodes, then extracting the dynamics state and global spectrum characteristics of the nodes, constructing a basic characteristic vector, utilizing an improved double-current interaction supergraph neural network, converging key information in the superside through local attention flow, capturing the total situation length Cheng Yilai by utilizing multi-scale topological diffusion flow in parallel, finally introducing a gating residual error fusion mechanism to self-adaptively integrate double-current characteristics, and accurately outputting propagation source probability based on a weighted cross entropy loss function. The method can effectively overcome the path reconstruction deviation of the traditional method and improve the robustness in the aspect of processing high-order interaction and long-distance dependence. The invention breaks through the performance bottleneck based on simple graph tracing and provides powerful technical support for positioning the propagation source.

Inventors

• ZHAN XIUXIU
• ZHANG CHENGJUN
• KE QIAO
• LIU CHUANG

Assignees

• 杭州师范大学

Dates

Publication Date

20260505

Application Date

20260128

Claims (4)

1. 1. A hypergraph propagation source positioning method based on interaction enhancement is characterized by constructing a hypergraph propagation structure for describing group interaction, acquiring node state observation in a propagation process, generating low-dimensional feature representation composed of observation states and global spectrum position information for each node, reducing feature redundancy and retaining key topology information, introducing an interaction enhancement feature cooperation mechanism into a hypergraph neural network, modeling local propagation features in a hyperedge and global topology features of a network layer in parallel, forming unified node representation in a self-adaptive fusion mode, training a model based on a weighted optimization target, outputting probability results of the nodes becoming propagation sources, and realizing effective positioning of the propagation sources.
2. 2. The method for locating a propagation source of a hypergraph based on interaction enhancement as claimed in claim 1, wherein the method comprises the following steps: step (1) constructing a hypergraph, spreading and acquiring an infection snapshot; (1-1) establishing a hypergraph Node set The superside set is N is the number of nodes, M is the number of supersides, and supersides Comprising any plurality of nodes for simulating a physical group contact event, ; (1-2) In hypergraph Randomly selecting a plurality of nodes as initial infection nodes, and adopting an SI infectious disease model for transmission; (1-3) at a certain time in the propagation process, observing the infection states of all nodes, and obtaining a complete infection snapshot Node infection state set Node infection status 1 Or 0,1 representing an infected state, 0 representing an uninfected state, ; Step (2) constructing basic feature vectors, namely constructing a super-graph Laplace matrix And performing feature decomposition to extract feature vector corresponding to minimum non-zero feature value as node Node spectral features of (2) Node spectral features Being able to capture the relative position of nodes in a hypergraph global topology, for each node Constructing a two-dimensional feature vector Including node infection status Sum node spectral features The two-dimensional feature vector As a node Input features of (a) ; Step (3) feature learning based on a dual-stream interaction mechanism; (3-1) input features to be input The hypergraph neural network model comprises two parallel processing flows, namely a hyperedge inner attention gathering flow and a multi-scale topological diffusion flow, and generates local gathering characteristics And global structural features ; (3-2) Double-stream interaction and residual fusion: merging local aggregation features And global structural features After splicing, calculating a fusion coefficient by using a sigmoid function through a linear layer mapping : Wherein, the method comprises the steps of, The splice is indicated as being a function of the splice, Representing linear layer mapping, then according to fusion coefficients The two paths of characteristics are weighted and summed, and the original input characteristics are superposed By means of Activating the function to obtain a final node representation : ; Step (4) source positioning prediction and optimization; (4-1) output prediction node representation Outputting probability fraction of each node as propagation source through full connection layer and activation function K nodes with highest probability scores are selected as prediction propagation sources; (4-2) model training Using output prediction results Using weighted binary Cross entropy loss As a sole optimization target; For example nodes To predict a node in the propagation source, then Otherwise ; And The weight coefficients of the positive and negative samples are respectively; (4-3) optimizing, namely updating network parameters by using an Adam optimizer until loss converges to obtain a final neural network model; step (5) positioning a hypergraph propagation source by utilizing a final neural network model, wherein the model input is a hypergraph And infection snapshot 。
3. 3. The method for locating a propagation source of a hypergraph based on interaction enhancement as claimed in claim 2, wherein the hypergraph laplacian matrix is constructed in the step (2) The method comprises the following steps: first, building an association matrix Incidence matrix Middle element Calculating a node degree matrix And a superside matrix Node degree To include nodes Is used for the number of the superedges of the (a), As an N diagonal matrix, diagonal elements Superlimit degree Is beyond the limit The number of nodes to be included is the number, For M M diagonal matrix, diagonal elements Normalization is carried out to obtain a hypergraph Laplacian matrix , Is a matrix of units which is a matrix of units, Representing the transpose.
4. 4. The interaction-enhancement-based hypergraph propagation source positioning method according to claim 2, wherein: The hyperedge internal attention gathering flow is characterized in that a hyperedge attention mechanism is designed aiming at the hypergraph local characteristics, and each hyperedge is designed The importance weights of all nodes in the superside are calculated by utilizing the multi-head attention, key nodes with larger contribution to transmission in the superside are dynamically focused, and local aggregation characteristics are generated ; The multi-scale topological diffusion flow is used for constructing primary and secondary power forms of hypergraph adjacent matrix And Simulating a multi-hop diffusion process propagated on the hypergraph, propagating node characteristics on topological structures with different scales, capturing long-distance dependency relationships, and generating global structural characteristics 。

Description

Hypergraph propagation source positioning method based on interaction enhancement Technical Field The invention belongs to the technical field of computer science and network security intersection, in particular relates to a hypergraph propagation source positioning method based on interaction enhancement, and belongs to a complex network propagation dynamics analysis and hypergraph neural network modeling method. According to the method, a hypergraph structure capable of describing high-order group interaction relations is constructed, an interaction enhanced feature learning mechanism is introduced, and the combined modeling is carried out on the structural information and the state information of the propagation process in the network, so that the high-precision positioning of source nodes of propagation events such as infectious diseases is realized. Background In modern society with high informatization and networking, various kinds of propagation behaviors frequently occur and rapidly evolve in complex network systems. The propagation process has the characteristics of strong burst, high diffusion speed, wide influence range and the like, and once the propagation process is out of control, serious social and economic losses can be caused. The problem of propagation source localization is becoming an important research topic in the fields of network science and security. The problem is typically to reverse infer the initial set of nodes that caused the propagation, given only the known network topology and the snapshot of the propagation state observations at a certain moment. However, the propagation process has randomness and irreversibility, and information deletion and noise interference are often accompanied in actual observation, and along with expansion of the propagation range, the structure and state characteristics left by the source node are continuously weakened, so that the traceability problem presents significant challenges such as huge solution space, high uncertainty and the like. Therefore, how to achieve fast and reliable propagation source localization under limited observation conditions has important practical significance for propagation control, risk intervention and decision making. The existing propagation tracing method is mostly built on a simple network model. Early studies often relied on heuristic metrics or optimization criteria such as rumor centrality, minimum descriptive length, etc., and in recent years deep learning models such as graph neural networks have been gradually introduced to mine propagation characteristics. However, such methods are generally based on the assumption that only binary pairwise interactions exist between nodes, and it is difficult to effectively describe the phenomenon of population interactions that is common in real-world propagation, such as in-home propagation, social group diffusion, and the like. When facing to a high-order interactive structure, forced disassembly of the group relationship into a common edge often leads to key information loss, thereby affecting the accuracy and stability of the tracing result. In contrast, the hypergraph model can naturally describe the group-level interaction relationship of 'one-to-many' or 'many-to-many' by allowing one hyperedge to be connected with a plurality of nodes at the same time, and provides a more reasonable structure expression form for describing complex propagation behaviors. Propagation traceability research is carried out under the hypergraph frame, so that higher-order interaction information is reserved from the structural level, and the positioning performance of multi-source propagation and complex propagation scenes is improved. With the development of geometric deep learning technology, the hypergraph neural network (HGNN) is gradually an important tool for processing high-order non-Euclidean structural data. Unlike traditional graph neural networks, HGNN can directly define messaging and feature aggregation mechanisms at the superside level, thereby capturing nonlinear high-order correlation patterns between multiple nodes. The hypergraph neural network is introduced into a propagation source positioning task, so that dependence on static indexes of artificial design can be eliminated, and propagation dynamics characteristics can be adaptively modeled in an end-to-end learning mode. However, how to effectively fuse local group information and global structural features under the hypergraph framework is still a problem to be solved in the prior art. Disclosure of Invention The invention aims to provide a hypergraph propagation source positioning method based on interaction enhancement, aiming at the problems that the structure expression capability is limited and the positioning result is easy to be interfered by noise in a high-order interaction scene of the existing propagation source positioning method. According to the invention, the hypergraph neural network is taken as