发表机构
Nanjing University of Information Science and Technology（南京信息工程大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文提出动态谱滤波（DSF），将时序图传播机制建模为随时间演化的切比雪夫多项式滤波器，在多个时序链接预测基准上取得优异性能，且参数、内存和训练时间远少于同类方法，适用于对计算效率要求高的场景。
AI 中文摘要
时序图学习通常围绕节点状态演化或交互历史编码展开，本文研究一个未被充分探索的、以算子为中心的问题：图传播机制本身是否应随时间演化？我们提出动态谱滤波（Dynamic Spectral Filtering, DSF），该方法在快照t时刻的传播由带向量值时变系数的切比雪夫多项式滤波器表示，将这些紧凑的多阶系数显式作为循环时序状态处理。一个循环分支提出更新，而乘法型全局门和阶专属门调节其幅度，该时序状态与节点数量无关。在MOOC、Wikipedia和Reddit时序链接预测基准上，收敛后的DSF分别取得0.7851、0.9088和0.9860的平均精度（AP）分数，可训练参数为93K至133K，峰值GPU内存为68至182 MB，每轮训练时间为1.6至2.1秒。与密切相关的DEFT基线相比，DSF在MOOC上表现更优，在Reddit上AP差距小于0.001，在Wikipedia上略低，同时使用的参数少8.3至8.6倍，GPU内存少25至33倍，每轮时间少5至19倍；相对于所有被测替代方法，其GPU内存少3.3至38.6倍。这些结果表明，当计算效率是首要要求时，直接的谱响应演化是一种有用的时序归纳偏置。
英文摘要
Temporal graph learning is commonly organized around the evolution of node states or the encoding of interaction histories. We study an underexplored, operator-centric question: should the graph propagation mechanism itself evolve over time? We introduce Dynamic Spectral Filtering (DSF), which represents propagation at snapshot t by a Chebyshev polynomial filter with vector-valued, time-dependent coefficients. DSF explicitly treats these compact multi-order coefficients as recurrent temporal states. A recurrent branch proposes updates, while multiplicative global and order-specific gates regulate their magnitude. The temporal state is independent of the number of nodes. On MOOC, Wikipedia, and Reddit temporal link-prediction benchmarks, converged DSF runs attain AP scores of 0.7851, 0.9088, and 0.9860, respectively, with 93K to 133K trainable parameters, 68 to 182 MB peak GPU memory, and 1.6 to 2.1 seconds of training per epoch. Against the closely related DEFT baseline, DSF is better on MOOC, within 0.001 AP on Reddit, and modestly lower on Wikipedia, while using 8.3 to 8.6 times fewer parameters, 25 to 33 times less GPU memory, and 5 to 19 times less time per epoch. Relative to all measured alternatives, it uses 3.3 to 38.6 times less GPU memory. These results support direct spectral-response evolution as a useful temporal inductive bias when computational efficiency is a first-class requirement.
CommentsCode is available at: https://github.com/YKong2018/DSF4TGL
