Enhancement of the Efficiency of Human Activity Recognition Under Severe Label Scarcity Using Adaptive and Hybrid Augmentation Strategies

Abstract

Large labelled datasets are required for Human Activity Recognition models. However, collecting and labeling ample amounts of data is expensive, time-consuming, and often impossible in many practical applications, including wearable health monitoring, elder care, and mobile sensing. Data augmentation is an effective strategy against label sparsity by artificially augmenting the training distributions, although its effectiveness exhibits wide variations across different datasets and model designs. This research assesses 15 different augmentation methods applied to UCI HAR, WISDM, and PAMAP2 datasets with different levels of annotated data ranging from 1% to 20% and explores the performance under the condition of limited labels. The findings reveal that the combination of jittering (Gaussian perturbation) and time warping (smooth temporal distortion) is the most consistent and reliable performing method across all three datasets. Interestingly, this combination of techniques not only surpassed individual augmentations and baseline models with no augmentation, but also provided substantial increases in accuracy even when a maximum of 5% labelled data is available.