Smartphone activity recognition improves with combined domain generalization methods
Domain Generalization for Smartphone-Based Human Activity Recognition: A Systematic Analysis of Components and Interactions
Artificial Intelligence
Summary
Smartphone apps that recognize human activities work less well when factors like different users or phone positions change. The paper studies methods to make these apps work better across such changes without seeing new data in advance. The authors tested many ways to train and design models and found that combining multiple techniques together usually works better than using any single method alone. They also found current ways to pick the best model during training miss large improvements that could be possible.
What this means in practice
- •For mobile app developers: Build smartphone apps that accurately detect activities even when users, devices, or phone placements vary, improving model reliability in real-world settings.
- •For wearable device engineers: Design wearable activity monitors robust to different users and environments by integrating combined domain generalization techniques for better cross-device performance.
Authors
Otávio Oliveira Napoli, Edson Borin
Abstract
Smartphone-based Human Activity Recognition (HAR) models often degrade under distribution shifts caused by changes in users, devices, sensor placements, environments, and acquisition protocols. Domain Generalization (DG) addresses this problem by learning from source domains without access to target data. Existing DG methods span training objectives, representation initialization, and architectural modifications, but these components are typically evaluated in isolation despite operating at different stages of the learning pipeline. We present a large-scale controlled benchmark of DG for smartphone-based HAR, comprising more than 410,000 experiments across four model architectures, thirteen training objectives including Empirical Risk Minimization (ERM), five initialization strategies, four architectural configurations, and two shift scenarios: cross-dataset and cross-position. Results show that individual DG components provide limited and highly conditional gains. Alternative objectives rarely outperform ERM consistently, self-supervised initialization helps in specific settings, and architectural modifications, particularly Dynamic Domain Generalization, provide the clearest standalone improvements. Joint configurations, however, frequently outperform their individual components and exhibit complementary and sometimes super-additive interactions, although gains remain model- and shift-dependent. Class-level analysis shows that the strongest configurations mainly improve difficult, shift-sensitive decision boundaries. Finally, oracle checkpoint analysis reveals substantial unrealized performance: source-validation selection recovers only 53% and 26% of the available oracle gain in cross-dataset and cross-position settings, respectively. Overall, effective HAR domain generalization requires jointly designing DG components and robust model-selection strategies.