📝 Publications

TS-Memory: Plug-and-Play Memory for Time Series Foundation Models
Sisuo Lyu, Siru Zhong, Tiegang Chen, Weilin Ruan, Qingxiang Liu, Taiqiang Lv, Qingsong Wen, Raymond Chi-Wing Wong, Yuxuan Liang
[Paper]  [Code]

• This work proposes TS-Memory, a plug-and-play memory adapter for time series foundation models that compiles retrieval-induced distributional knowledge from an offline, leakage-safe kNN teacher into a lightweight parametric module. By freezing the TSFM backbone and applying confidence-gated memory distillation, TS-Memory enables retrieval-free, constant-time inference while improving both point and probabilistic forecasting across diverse TSFM backbones and benchmarks.

OccamVTS: Distilling Vision Models to 1% Parameters for Time Series Forecasting
Sisuo Lyu, Siru Zhong, Weilin Ruan, Qingxiang Liu, Qingsong Wen, Hui Xiong, Yuxuan Liang
[Paper]  [Code]  [Poster]

• This work proposes OccamVTS, a novel knowledge distillation framework that extracts only the essential 1% of predictive information from large vision models (LVMs) to address the severe parameter redundancy and semantic misalignment challenges in time series forecasting by employing pre-trained LVMs as privileged teachers to guide lightweight student networks through pyramid-style feature alignment and selective distillation, effectively filtering out irrelevant high-level semantic noise while preserving crucial low-level temporal patterns.

Improving Foundation Model Group Robustness with Auxiliary Sentence Embeddings
Sisuo Lyu, Hong Liu, Jie Li, Yan Teng, Yingchun Wang
[Paper]  [Code]

• This work proposes DoubleCCA, a novel framework that leverages auxiliary sentence embedding models to enhance group robustness of vision-language foundation models against spurious correlations, addressing critical bias mitigation challenges in trustworthy AI deployment through a two-stage canonical correlation analysis technique that first aligns augmented and original text embeddings in a shared semantic space, then reconstructs invariant features to merge with visual representations, effectively reducing sensitivity to group-based biases.

Multi-view Hypergraph-based Contrastive Learning Model for Cold-Start Micro-video Recommendation (Oral)
Sisuo Lyu, Xiuze Zhou, Xuming Hu
[Paper]  [Code]  [PPT]  [Poster]

• This work proposes MHCR, the first model to leverage hypergraphs and contrastive learning for cold-start micro-video recommendation, implementing multi-view multimodal feature extraction with user-item graph, item-item affinity graph, and hypergraph layers combined with cross-modal and graph-hypergraph contrastive learning to address sparse interaction signals and over-smoothing challenges in platforms like TikTok and Kwai.