ProtoFlow: Prototype-Guided Flow Matching for Multivariate Time Series Forecasting
ProtoFlow forecasts multivariate time series by flow-matching from learned vector-quantization prototypes instead of Gaussian noise.
ProtoFlow is a multivariate time-series forecasting framework that combines a vector-quantized autoencoder with flow matching. Instead of starting transport from Gaussian noise, it builds a structured prior from prototypes in the trained codebook. A Diffusion Transformer rectified flow then maps those prototypes to future latent representations conditioned on history. The authors report better benchmark forecasting than autoregressive token methods, without their exposure bias, and with efficient non-iterative inference.
- ProtoFlow combines vector-quantized autoencoding with prototype-prior flow matching.
- The learned codebook replaces a generic Gaussian flow prior.
- A DiT rectified flow predicts future latents from historical observations.
- The method avoids autoregressive rollout mismatch and reports efficient inference.
- AI models
- ProtoFlow
From the original article
1 min read at arxiv.orgGenerative modeling has shown strong promise for multivariate time mseries (MTS) forecasting, especially scale to high-dimensional settings. Diffusion-based methods achieve competitive performance but typically require many sampling steps at inference. VAE-based non-iterative forecasting frameworks have therefore emerged as an efficient alternative.
By Shibo Feng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.