WaveStitch: Flexible and Fast Conditional Time Series Generation with Diffusion Models

TL;DR

WaveStitch tackles conditional time-series generation with dual conditioning on metadata and observed signals, addressing limitations of prior methods in conditioning scope, generalization, and coherence. It combines a hybrid training-inference approach with a pipelined parallel architecture and a stitching mechanism to maintain temporal consistency across overlapping windows. Empirically, it achieves a 1.81x reduction in mean-squared-error on average and up to 166.48x faster generation than autoregressive methods while preserving coherence, across five diverse datasets. The approach enhances practicality for forecasting, imputation, and generation under varied missingness patterns and metadata configurations, with open-source code available for reuse.

Abstract

Generating temporal data under conditions is crucial for forecasting, imputation, and generative tasks. Such data often has metadata and partially observed signals that jointly influence the generated values. However, existing methods face three key limitations: (1) they condition on either the metadata or observed values, but rarely both together; (2) they adopt either training-time approaches that fail to generalize to unseen scenarios, or inference-time approaches that ignore metadata; and (3) they suffer from trade-offs between generation speed and temporal coherence across time windows--choosing either slow but coherent autoregressive methods or fast but incoherent parallel ones. We propose WaveStitch, a novel diffusion-based method to overcome these hurdles through: (1) dual-sourced conditioning on both metadata and partially observed signals; (2) a hybrid training-inference architecture, incorporating metadata during training and observations at inference via gradient-based guidance; and (3) a novel pipeline-style paradigm that generates time windows in parallel while preserving coherence through an inference-time conditional loss and a stitching mechanism. Across diverse datasets, WaveStitch demonstrates adaptability to arbitrary patterns of observed signals, achieving 1.81x lower mean-squared-error compared to the state-of-the-art, and generates data up to 166.48x faster than autoregressive methods while maintaining coherence. Our code is available at: https://github.com/adis98/WaveStitch

Figures (2)

• Figure 6: Parallel Inference Conditioning with Stitching and Self-Guidance Loss. The denoising process flows from left to right. At each step $t$, the model receives denoised windows from step $t+1$. It then refines these using two types of gradient corrections: the self-guidance loss (in blue), computed using observed signals, and the stitching loss (in orange) on the overlaps. These corrections update the metadata-conditioned estimate before passing it to the next denoising step $t-1$.
• Figure 7: Autocorrelation (up to 100 lags), and Cross-feature Correlations for BQ for R, I, and B tasks.