Adaptive multi-scale hypergraph neural network for multi-source detection

doi:10.1016/j.knosys.2026.116883

Adaptive multi-scale hypergraph neural network for multi-source detection

2026-08-01

0

PRE

AI

DOI：10.1016/j.knosys.2026.116883

Original

Original request for help

Save

Abstract

En 中文

Hypergraph-based source detection focuses on inversely identifying diffusion sources from spreading snapshots defined over higher-order interaction structures. Traditional methods developed on simple networks assume pairwise transmission and therefore fail to capture the higher-order aggregated interactions and non-binary dependencies that commonly exist in real-world systems. Since hypergraphs explicitly encode multi-node relationships, investigating source detection under hypergraph representations is essential for improving spreading control and early warning in complex environments. To address the multi-source tracing task characterized by higher-order features, we propose an Adaptive Multi-Scale Hypergraph Neural Network framework, termed AMHNN. The proposed AMHNN incorporates an adaptive spreading layer that integrates learnable gating with multi-order topological perception, enabling the model to alleviate source signal dilution, namely the loss of source-related information caused by repeated aggregation with infected non-source nodes, while effectively capturing long-range structural dependencies. In addition, a hyperedge-based attention mechanism is designed to model local heterogeneity by evaluating node importance strictly within hyperedge scopes. This unified architecture facilitates robust learning of latent diffusion patterns directly from observed data without requiring prior knowledge of spreading dynamics. Experiments on 12 synthetic and empirical hypergraph datasets demonstrate that AMHNN consistently outperforms baseline methods, achieving an average F1-score improvement of 13.8% over the strongest baseline across the three spreading models. The proposed framework provides an effective and scalable solution for source detection in higher-order networks and supports practical applications in epidemic tracking, rumor identification, and complex system governance.

Keywords:

Hypergraph source detection

Hypergraph neural network

Adaptive gating

Hyperedge attention

Adaptive multi-scale hypergraph neural network for multi-source detection

Abstract

Journal

Organization

Cited Papers