EFDiff: Frequency-informed Diffusion for Extreme-value Time Series Generation

Lin Jiang, Rongchao Xu, Dahai Yu, Guang Wang

Keywords: Time-series Generation, Extreme Event, Diffusion Model, Frequency Domain

TL;DR: We design a frequency-informed extreme-aware time-series generation framework

Abstract:

Time-series generation, which aims to produce realistic synthetic sequences that preserve temporal dynamics, is essential for data augmentation and practical applications. However, existing methods often fail to capture extreme-value distributions, which are crucial in domains such as finance, climate, and energy. This limitation mainly stems from overall-fit objectives and smoothing procedures that distort extreme-event structures. To address these challenges, we propose EFDiff, a frequency-informed extreme-aware time-series generation framework. Unlike conventional approaches that focus on long-tail preservation in the time domain, EFDiff adopts a frequency-domain perspective by integrating a frequency-based disentanglement strategy into diffusion models. The key innovation lies in an Extreme Component, which consists of two key modules: (i) Extreme-Frequency Extraction (EFX), which constructs a global extreme-frequency dictionary that characterizes potential extreme patterns via event-driven local analysis and multi-metric integration based on the proposed concept of extreme-contributing frequencies; and (ii) Extreme-Frequency Generation Enhancement (EFGEN), which includes a novel Transformer-based Soft Frequency Selection Network to identify relevant frequencies and effectively model extreme patterns during
the denoising process. Extensive experiments on five real-world datasets across six evaluation metrics demonstrate that EFDiff consistently achieves strong overall generation quality and substantially improves the fidelity of extreme-value generation.

Primary Area: learning on time series and dynamical systems

Submission Number: 7972