Directional Time Series Editing via Retrieval-Guided Jacobian-Vector Inference
TL;DR
A new TSE setting for continuous, magnitude-aware condition transitions is introduced and JAVELIN, a retrieval-guided framework for directional editing via JAcobian-VEctor Latent INference is proposed, enabling precise, content-preserving edits without retraining the generative model.
Abstract
Time Series Editing (TSE) synthesizes realistic time series by modifying existing trajectories under user-specified conditions, with growing importance across many applications. However, existing TSE formulations are largely restricted to discrete or categorical controls and struggle to handle continuous condition shifts, especially at large magnitudes, in a unified and robust way. To address this limitation, we introduce a new TSE setting for continuous, magnitude-aware condition transitions and propose JAVELIN, a retrieval-guided framework for directional editing via JAcobian-VEctor Latent INference. JAVELIN constructs a query-specific local neighborhood and learns a lightweight latent editor at inference time, enabling precise, content-preserving edits without retraining the generative model. Extensive experiments on synthetic and real-world datasets show that JAVELIN produces high-fidelity, condition-coherent, and controllable edits, substantially outperforming existing generation and editing baselines. The source code is available at https://github.com/AmethystQ/JAVELIN/.