Domain Generalization for Smartphone-Based Human Activity Recognition: A Systematic Analysis of Components and Interactions
| Source: arXiv AI
Tags: domain generalization, human activity recognition, smartphone sensors, benchmark, transfer learning
A large-scale benchmark of 410,000+ experiments on smartphone-based human activity recognition finds that individual domain generalization techniques rarely beat ERM, but joint configurations frequently outperform their parts — sometimes super-additively — while checkpoint selection recovers only 26–53% of available oracle gain.
Details
Napoli and Borin present the most systematic study to date of domain generalization (DG) for smartphone-based human activity recognition. HAR models degrade when users, devices, sensor placements, or environments change; DG aims to learn representations that transfer across these shifts. The benchmark covers four model architectures, thirteen training objectives including ERM as baseline, five initialization strategies, four architectural configurations, and two shift scenarios (cross-dataset and cross-position) — totaling over 410,000 experiments. Key findings: alternative training objectives rarely outperform ERM consistently; self-supervised initialization helps only in specific settings; Dynamic Domain Generalization provides the clearest standalone architectural improvement. However, joint configurations frequently exhibit complementary or super-additive interactions. A key practical finding is that checkpoint selection is severely underestimated as a bottleneck: source-validation selection recovers only 53% of oracle gain in cross-dataset settings and only 26% in cross-position. This suggests model selection strategy may matter as much as the DG method. The work spans a total of 27 pages with 14 figures and 10 tables, making it a comprehensive reference for HAR researchers.