Reproduce “Harnessing Vision Models for Time Series Analysis: A Survey” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingchao Ni, Ziming Zhao, ChengAo Shen, Hanghang Tong, Dongjin Song, Wei Cheng, Dongsheng Luo, Haifeng Chen
The survey presents a framework for turning time series into images for vision-model tasks. It covers image encodings. In a zero-shot study across six datasets and four horizons, VisionTS reported average mean squared error of 0.309, versus 0.327 for MOIRAI Small and 0.310 for MOIRAI Base. The survey did not run this experiment, and the best encoding remains open.
Key insights
- Vision models, particularly LVMs and LMMs, offer inherent advantages over LLMs for time series analysis due to their alignment with continuous data, natural encoding of multivariate correlations, and potential for compact representation of long time series.
- A structured, dual-view taxonomy is crucial for understanding the diverse methodologies involved in converting time series into visual formats and subsequently processing them with various vision architectures.
- Successful application of vision models for time series requires careful attention to critical pre-processing steps like time series normalization and image alignment, as well as post-processing for accurate time series recovery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.