摘要
文章提出了一种创新的时序异构图自监督嵌入框架,有效捕获图数据的异构性与时序动态特性。实验表明,该方法在链接预测和节点分类等任务中表现卓越,较现有基准方法平均提升1.7%~2.3%的性能。该框架优势源于三个核心创新:1)层次化编码机制高效处理多类型节点与边的复杂交互;2)基于Transformer的时序自注意力模型精确捕获长短期依赖关系;3)融合掩码重建、对比学习与结构保持的多目标自监督框架显著减轻了对数据标注的依赖。消融研究证实,时序信息建模与异构关系建模对性能贡献最为显著,分别带来2.9%和2.3%的提升,同时文章提出的方法在处理稀疏关系方面表现也较为突出。本研究不仅推进了时序异构图表示学习的理论前沿,也为社交媒体分析、电商推荐等现实应用提供了有效的解决方案。
This paper proposes an innovative temporal heterogeneous graph self-supervised embedding framework that effectively captures both graph heterogeneity and temporal dynamics.Experi-ments demonstrate that the proposed method excels in tasks such as link prediction and node classification,achieving 1.7%~2.3%average performance improvements over existing baseline methods.The framework’s advantages stem from three key innovations:1)a hierarchical encod-ing mechanism that efficiently processes complex interactions between diverse node and edge types;2)a Transformer-based temporal self-attention model that precisely captures both longterm and short-term dependencies;and 3)a multi-objective self-supervised learning framework integrating mask reconstruction,contrastive learning,and structure preservation that signifi-cantly reduces reliance on data annotation.Ablation studies confirm that temporal information modeling and heterogeneous relationship modeling contribute most significantly to performance gains,bringing improvements of 2.9%and 2.3%,respectively.Meanwhile,the proposed method also demonstrates outstanding performance in handling sparse relationships.This research not only advances the theoretical frontier of temporal heterogeneous graph representation learning but also provides effective solutions for real-world applications such as social media analysis and e-commerce recommendations.
作者
建一飞
Yifei Jian(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai)
出处
《建模与仿真》
2025年第5期644-658,共15页
Modeling and Simulation
关键词
时序异构图
自监督学习
图神经网络
表示学习
Temporal Heterogeneous Graph
Self-Supervised Learning
Graph Neural Network
Representation Learning
