Reproduce “Out-of-Distribution Representation Learning for Time Series Classification” on OpenResearch

About this paper

Wang Lu, Jindong Wang, Xinwei Sun, Yiqiang Chen, Xing Xie

DIVERSIFY, developed by researchers from ICT CAS, Microsoft Research Asia, and Fudan University, proposes an out-of-distribution representation learning algorithm for time series classification that automatically characterizes unknown latent distributions within training data. It consistently achieves superior generalization performance across various tasks, including a 4.3 percentage point increase in gesture recognition accuracy and a 9.9 percentage point improvement in cross-dataset human activity recognition.

Key insights

• Time series data contains inherent, dynamic latent distributions (temporal and spatial shifts) that can be discovered and leveraged for robust OOD generalization without explicit domain labels.
• Maximizing the divergence among identified latent sub-domains within the source data, followed by learning domain-invariant features, is a theoretically supported strategy for improving generalization to truly unseen target distributions.
• The use of class-invariant features during pseudo-domain assignment and the integration of 'pseudo domain-class labels' are critical for effective disentanglement of class and domain information.

Summary by alphaXiv.