计算机科学
图形
人工智能
残余物
融合
同种类的
节点(物理)
语义学(计算机科学)
传感器融合
混乱
人工神经网络
机器学习
语义解释
语义记忆
融合机制
语义网络
知识图
语义计算
语义相似性
深层神经网络
GSM演进的增强数据速率
语义属性
数据挖掘
语义数据模型
异构网络
作者
Pengfei Li,Qi Ye,Haitian Huang,Zhenyu Wang
出处
期刊:Conference proceedings
[Institute of Electrical and Electronics Engineers]
日期:2025-10-05
卷期号:: 1775-1782
标识
DOI:10.1109/smc58881.2025.11343687
摘要
Heterogeneous graphs(HGs), characterized by diverse node and edge types as well as complex semantics, pose significant challenges for graph neural networks. Existing methods suffer from three major limitations: (1) homogeneous aggregation leads to semantic confusion and weakens local structural information; (2) the importance of different semantic paths is difficult to distinguish due to a lack of dynamic semantic fusion mechanisms; (3) single-view modeling fails to capture cross-view interactions between structural and semantic information. To address these issues, we propose MHG-DVA, a multi-scale dual-view attention model framework that integrates both the structural view and the meta-path view. The structural view employs gated attention and structural-level fusion to finely model heterogeneous neighbors, while the metapath view combines intra-path semantic propagation with interpath context-aware attention to capture multi-hop semantic dependencies. A node-adaptive fusion module is further introduced to dynamically integrate multi-view information through gated residual connections and cross-view interactions. Experiments on DBLP, ACM, Yelp, and Amazon show that MHG-DVA outperforms state-of-the-art methods in node classification (+0.95%, +2.59%, +0.48% Micro-F1) and fraud detection, where it surpasses IDGL by +0.89% and +1.03% in AUC under 20% and 40% training settings, respectively. Ablation studies confirm the efficacy of the multi-scale mechanism, establishing a novel fusion paradigm for heterogeneous graph analysis with both theoretical and practical value.
科研通智能强力驱动
Strongly Powered by AbleSci AI