Reproduce “Feature Anchors for Time-Series Sensor-Based Human Activity Recognition” on OpenResearch

About this paper

Ruijie Yao, Chenhang Li, Danyang Zhuo, Tingjun Chen, Xiaoyue Ni

The Temporal Conditioning Network (TCNet) introduces a method to refine explicit handcrafted time-series features (TSFs) using neural context derived from raw sensor signals. This approach yields competitive or superior performance on five Human Activity Recognition (HAR) benchmarks with substantially fewer parameters, while maintaining the interpretability of features.

Key insights

• Treating handcrafted features as adaptable anchors, refined by neural context, provides a more efficient and interpretable representation compared to purely latent deep models or static feature pipelines.
• The anchor correction mechanism, which modulates explicit TSFs based on raw-signal context, is critical for TCNet's performance gains, rather than merely combining multiple feature types.
• Deep networks, even when task-trained, often weakly encode or suppress the linear decodability of important handcrafted TSFs, highlighting the value of explicit anchor retention.
• Lightweight, in-domain, anchor-aligned pretraining can be more effective for structured HAR transfer learning under domain shift than large-scale, cross-domain latent pretraining.

Summary by alphaXiv.