ADL-GAN: Data Augmentation to Improve In-the-Wild ADL Recognition Using GANs

The types of Activities Daily Living (ADL) a person performs or avoids, and underlying patterns can provide insights into physical mental health, making passive ADL recognition from smartphone sensor data important. However, as people perform ADLs unequally in real life, datasets collected the wild be extremely imbalanced, which presents challenge to Machine Learning (ML) classification. Prior solutions mitigating imbalance, such oversampling instance weighting, reduce but do not completely eliminate problem. We instead propose ADL-GAN, utilizes translation Generative Adversarial Networks (GANs), synthesize motion audio improve classification performance. ADL-GANs augment minority subject $A$ by translating samples either 1) other where has adequate Context-transfer ADL-GAN 2) subjects with Subject-transfer ADL-GAN. utilize multi-domain contrastive loss functions many-to-many translations between classes subjects, respectively. outperformed baselines improved balanced accuracy (BA) on an in-the-wild dataset 27.9 %, while context-transfer performed best scripted dataset, improving BA 9.58 %. augmented were shown more realistic diverse than conditional GAN.