CV

Addressed the lack of cross-domain transferability in supervised spatio-temporal forecasting. Designed a self-supervised pretraining framework combining temporal masking and a efficient spatial-temporal encoder. Compared with concurrent SOTA supervised baselines (GPT-ST), ST-ReP achieved achieved 20.24% lower MAE, 67.22% lower GPU usage and 28.07% lower training time. Contribution: Pioneered self-supervised spatio-temporal representation pretraining, bridging efficient forecasting and future foundation-model paradigms.*

Solved the inefficiency of Transformers and the neglect of propagation time-lag among sensors. Introduced Time-Lag Embedding and Spatial Proxy Attention (SPA) to capture asynchronous correlations with O(N) complexity. Compared with the SOTA model at publication time (STAEformer), TLAST achieved 2.81% lower MAE, 92.50% lower GPU usage and 91.32% lower training time. Contribution: Established a scalable Transformer paradigm that unifies lag-aware modeling with linear complexity.

Tackled the limitation that prior GNN-based models neglect explicit temporal priors and cross-time spatial correlations. Proposed a time-encoding module and a time-adjoint graph convolution capturing multi-scale temporal-spatial semantics with a concise architecture. Compared with the SOTA model at that time (ASTGNN), TAGnn achieved 9.27% lower MAE and 89.84% lower training time. Contribution: Demonstrated that explicit temporal priors and cross-time adjacency enable interpretable, lightweight, and accurate forecasting.

Addressed the challenge of dynamic and uncertain spatial dependencies in traffic forecasting. Proposed a fuzzy-neural dynamic adjacency generator to construct cross-time graphs and capture evolving spatial-temporal relations using a single GCN layer. Compared with the SOTA method at publication time (STSGCN), DSTAGCN achieved 8.71% lower MAE and 50.47% lower training time. Contribution: Introduced uncertainty-aware, time-varying adjacency learning, realizing both higher predictive accuracy and training efficiency.