Research map / Time series / Causal and generative time series

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

NeurIPS 2026 · Time series › Causal and generative time series · approach: Diffusion models for time series · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes E4GEN, an event-level explainable diffusion framework that activates a learned extreme-control signal late in denoising, predicts per-sample extreme-event semantics, and injects them via a gated control network to improve extreme-event time-series generation.

Problem

Existing time-series generators emphasize overall distributional fidelity and attenuate sparse extreme events; value-level and heavy-tailed methods do not model event-level structure (location, duration, shape).

Approach

Treat extreme events as coherent temporal events represented by location, intensity, and keypoint shape; learn when to activate control (E-Activator via Backbone Drift and Outlier Emergence Score), predict per-sample event semantics from the intermediate denoising state (SDSP, trained with DCT-NIS), and inject them into an x-prediction Diffusion-TS-style backbone through a gated Extreme Control Ne

Key results

• On Wea-Temp, E4GEN attains the best Wass. (0.9346 vs 1.0998 for HeavyDiff) and best EC-Diff (1090 vs 2587 for Diffusion-TS) and ED-W1 (0.2916 vs 0.4235).
• Across 102 dataset-metric comparisons, E4GEN reports state-of-the-art in 79.4% of cases and top-two in 95.1%.

Surprising findings

• Heavy-tailed value-level generation can amplify dominant event types while suppressing minor ones (a 'Matthew effect'), so heavier tails do not imply higher extreme-event fidelity.

Limitations

• Evaluated on only six datasets across four domains; generality beyond these is untested.
• Extreme-event definition depends on domain-specific thresholds, and generated semantics depend on threshold quality.
• Additional semantic prediction and control modules increase training and sampling cost.

Experimental setup

Datasets: Wea-Temp Wea-Prec LTST-ECG HH-Power PEMS-SF Synthetic dataset

Benchmarks: Wea-Temp Wea-Prec LTST-ECG HH-Power PEMS-SF Synthetic

Metrics: Wasserstein distance KS JS MMD ACD Context-FID Predictive score Reconstruction score EMean-W1 ECount-Diff

Baselines: TimeGAN TimeVAE koVAE F-Flow DiffWave Diffusion-TS SDForger FIDE HeavyDiff

Builds on: Diffusion-TS DDPM HeavyDiff FIDE

Closest papers by meaning

• FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation (NeurIPS 2024 · similarity 0.88)
• Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.86)
• TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation (ICML 2025 · similarity 0.85)
• Interacting Diffusion Processes for Event Sequence Forecasting (ICML 2024 · similarity 0.84)
• Diffusion-TS: Interpretable Diffusion for General Time Series Generation (ICLR 2024 · similarity 0.84)
• TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop (NeurIPS 2025 · similarity 0.84)