Reproduce “UniMTS: Unified Pre-training for Motion Time Series” on OpenResearch

About this paper

Xiyuan Zhang, Diyan Teng, Ranak Roy Chowdhury, Shuheng Li, Dezhi Hong, Rajesh K. Gupta, Jingbo Shang

UniMTS, developed by researchers from UC San Diego, Qualcomm, and Amazon, introduces the first unified pre-training framework for motion time series, enabling robust generalization across device locations, orientations, and activity types. It achieves an average 340% macro-F1 improvement in zero-shot classification over baselines and significantly enhances few-shot and full-shot performance on 18 real-world datasets, while also being more efficient.

Key insights

• Unified pre-training on simulated motion data can effectively generalize across device latent factors and activity semantics, reducing reliance on extensive real-world labeled datasets.
• Explicitly modeling spatio-temporal relationships with a graph encoder and leveraging semantically rich text descriptions from LLMs are crucial for robust cross-dataset generalization in motion time series.
• Achieving efficient pre-trained models suitable for edge device deployment is possible by optimizing architecture (e.g., graph encoder vs. CNN) and pre-training on synthetic data.

Summary by alphaXiv.