| 摘要: |
| 动态图表示学习通过捕捉实体间随时间演化的拓扑结构与交互模式, 为链接预测等下游任务提供具有时空感知能力的嵌入表示, 从而揭示复杂系统的动态演化规律. 连续时间动态图因其具备丰富的细粒度时间信息, 为社交网络演化等复杂过程提供了更贴近真实世界的建模范式. 然而, 当前连续时间动态图表示学习至少面临3个挑战: (1)在长交互序列中有效提取关键信息需具备强大的长期时间依赖建模能力; (2)处理长交互序列时需控制计算复杂度以提升计算效率; (3)传统按交互时间顺序捕获交互模式的方法难以揭示隐藏的非因果性关联. 针对上述问题, 提出一种连续时间动态图表示学习模型, 通过双向选择性状态空间编码机制, 既可捕捉交互序列的长期时间依赖, 又能通过反向路径引入后续交互信息, 打破时间单向性约束, 从而增强模型对全局上下文的理解能力. 大量实验结果表明, 该模型在不同领域真实世界数据集上表现出的预测性能均显著优于基线方法, 同时具有高效的计算效率, 使在有限计算资源下实现长时间双向依赖的建模成为可能. |
| 关键词: 动态图表示学习 连续时间动态图 双向选择性状态空间模型 长期双向依赖关系 |
| DOI：10.13328/j.cnki.jos.007654 |
| 分类号: |
| 基金项目:江苏省自然科学基金(20201430); 江苏省高校优秀中青年教师和校长海外研究培训计划; 江苏省政府海外留学奖学金; 江苏省研究生科研与实践创新项目(SJCX25_2293) |
|
| Dynamic Graph Representation Learning Based on Bidirectional Selective State Space Model |
|
HE Ping, WANG Lei-Da, XU Xiao-Hua, YAN Yong-Lin
|
|
College of Information and Artificial Intelligence (College of Industrial Software), Yangzhou University, Yangzhou 225127, China
|
| Abstract: |
| Dynamic graph representation learning captures time-evolving topological structures and interaction patterns among entities, providing spatiotemporally aware embedding for downstream tasks such as link prediction, revealing the dynamic evolution patterns of complex systems. Continuous-time dynamic graphs, owing to their rich fine-grained temporal information, provide a more realistic modeling paradigm for complex processes such as social network evolution. However, current continuous-time dynamic graph representation learning faces at least three challenges. (1) The effective extraction of critical information from long historical interaction sequences requires robust modeling of long-term temporal dependencies. (2) Handling long interaction sequences requires controlling computational complexity to improve efficiency. (3) Traditional approaches that capture interaction patterns in chronological order struggle to reveal hidden non-causal associations. To address these challenges, this study proposes a novel continuous-time dynamic graph representation learning model. By leveraging a bidirectional selective state-space encoding mechanism, long-term temporal dependencies in interaction sequences can be captured, and subsequent interaction information can also be incorporated through backward pathways. The temporal unidirectional constraint is thus broken, enhancing the model’s global context comprehension capability. Extensive experimental results demonstrate that the model consistently outperforms baseline methods in predictive performance on real-world datasets spanning diverse domains, while maintaining high computational efficiency, enabling the modeling of long-term bidirectional dependencies under limited computational resources. |
| Key words: dynamic graph representation learning continuous-time dynamic graph bidirectional selective state space model long-term bidirectional dependency |
