Dr. Guang Wang’s DCS Lab has one paper accepted at NeurIPS 2026

E4GEN: Event-level Explainable Extreme-Enhanced Time-series Generation

This paper is led by Dr. Wang’s Ph.D. student Lin Jiang, with all co-authors from the DCS Lab. The work addresses a critical challenge in time-series generation: while existing generative models can effectively capture overall data distributions, they often struggle to faithfully reproduce rare but consequential extreme events, which are particularly important in real-world applications such as climate, healthcare, energy, and transportation.

To address this challenge, the paper proposes E4GEN, an explainable diffusion framework for extreme-aware time-series generation. Rather than treating extreme-event generation as a black box, E4GEN systematically investigates when, what, and how to control extreme events throughout the generation process. It identifies when extreme-event control should be activated during diffusion, determines what extreme-event characteristics should be preserved for individual samples, and adaptively incorporates these characteristics into the denoising process, enabling more faithful generation of both overall temporal patterns and rare extreme events.

Evaluated on 6 real-world datasets spanning climate, healthcare, energy, and transportation using 17 evaluation metrics, E4GEN demonstrates strong performance across overall fidelity, extreme-event fidelity, and downstream utility compared with state-of-the-art time-series generation methods. These results highlight the importance of explicitly modeling extreme events in generative time-series models and provide a more controllable and interpretable approach to synthetic time-series generation for real-world systems where rare and high-impact events matter.