Time Series Analysis
Advancing time series forecasting and classification with modern deep learning, including self-attention analysis, sharpness-aware training, and data-driven anomaly detection.
Overview
Time series data is ubiquitous in finance, healthcare, and engineering. Our research develops principled deep learning methods tailored to the unique structure of sequential data.
Key Directions
Self-Attention for Time Series: We critically evaluate whether Transformer-based self-attention is truly beneficial for time series forecasting, challenging prevailing assumptions and proposing efficient alternatives (CATS, NeurIPS 2024).
Sharpness-Aware Training: We adapt sharpness-aware minimization (SAM) to exploit the periodic structure of time series, achieving faster convergence and better generalization in classification and forecasting tasks (Applied Soft Computing 2023).
Anomaly Detection: We combine differential privacy techniques with upsampling strategies to improve anomaly detection in highly imbalanced industrial datasets (Engineering Applications of AI 2026).
Diffusion-based Generation: We develop TimeBridge, a diffusion bridge model that uses learned priors for controllable and realistic time series generation (KDD 2026).