文献类型：期刊文献

英文题名：An Anomaly Detection Algorithm Based on Multi-Scale Features and Positional Decay in Graph Neural Networks

作者：Li, Ping[1]; Wang, Zhanquan[1]

机构：[1] School of Information Science and Engineering, East China University of Science and Technology, Shanghai, 200237, China

年份：2026

起止页码：2043

外文期刊名：2026 9th International Conference on Advanced Algorithms and Control Engineering, ICAACE 2026

收录：EI(收录号：20262420909578)

语种：英文

外文关键词：Decay (organic) - Graph algorithms - Graph neural networks - Message passing - Multilayer neural networks - Risk analysis - Risk assessment - Semantics - Signal detection

摘要：Anomaly detection on graphs plays a critical role in applications such as financial risk control, social network analysis, and cybersecurity. However, existing methods often struggle to capture multi-scale node semantics and suffer from over-smoothing when stacking multiple message-passing layers. To address these challenges, this paper proposes a novel Multi-Scale and Decay-Aware Graph Transformer Network (MSDAGTN). MSDA-GTN introduces a Progressive Scale Aggregation module to extract and fuse multi-scale node features, constructing hierarchical semantic representations. In addition, a Position-Aware Decay mechanism dynamically adjusts attention weights based on node-pair distances, enabling more effective long-range dependency modeling through multi-head attention. A gating-based feature fusion mechanism further improves robustness, and final anomaly prediction is conducted using a multi-layer perceptron. Experiments on the Yelp and Amazon datasets demonstrate that MSDA-GTN consistently outperforms baselines in terms of AUC, F1, and AP. Ablation studies also validate the effectiveness of each proposed component. ? 2026 IEEE.