CSFO: A Category-Specific Flattening Optimization Method for Sensor-Based Long-Tailed Activity Recognition

The sensor-based activity of daily living recognition (ADLR) has demonstrated exceptional performance across a wide array of applications. However, prevailing ADLR models predominantly concentrate on high-frequency activity categories, resulting in poor generalization when addressing long-tail categories that are underrepresented in the training data. To address this issue, we integrate a category-specific flattening optimization (CSFO) algorithm with the two-stage decoupling paradigm to enhance long-tail recognition (LTR) capabilities within the realm of sensor-based ADLR. In the first stage, the feature extractor and classifier are trained with class-conditioned parameter perturbations to improve resilience against local minima and enhance generalization. Specialized error functions adjust the perturbation scale based on each class’s sample distribution, constraining generalization error individually. In the second stage, the backbone network remains frozen while class-balanced sampling generates adversarial features to refine the classifier. The error function balances standard and adversarial losses, enhancing robustness and mitigating data imbalance impacts. Experiments on the UCI-HAR, OPPORTUNITY, WISDM, and USC-HAD datasets, all characterized by imbalanced distributions, demonstrated that the model improved accuracy by 0.66%, 0.28%, 0.98%, and 2.01%, respectively. Our experiments and analyses have demonstrated the exceptional robustness of the CSFO method in ADLR applications.