Research
Time-Series & Data Mining
We build robust learning for temporal and relational data — reversible normalization for accurate forecasting (RevIN), LLM-based and deep anomaly detection, graph learning, and recommendation — with an emphasis on staying reliable under distribution shift and noise.
28 publications
1,598 citations
Selected Publications — 5 most representative of 28
Pacer and Runner: Cooperative Learning Framework between Single- and Cross-Domain Sequential Recommendation
Revisiting LLMs as Zero-Shot Time Series Forecasters: Small Noise Can Break Large Models
Shows LLM zero-shot time-series forecasters are highly noise-sensitive and underperform simple models, arguing fine-tuning beats prompting for numerical sequences