Human Context Recognition: A Controllable GAN Approach

Human activity recognition (HAR), the use of sensor data to identify activities performed by an individual, has garnered substantial research attention. However, obtaining data for HAR is costly. As such, recent studies have used Generative Adversarial Networks (GANs) to create synthetic HAR data. Problematically, these solutions cannot generate personalized data that match the behavior of a user, limiting potential HAR applications as prior work has shown that personalized HAR models outperform generalized ones. To address this problem, we propose a novel controllable GAN where user and activity attributes can be explicitly requested for generation under one model. The GAN can learn these attributes independently and infer context pairs never seen during training: an impossibility for existing methodologies. We show that while our model is comparable to standard HAR GANs, its ability to consider user differences and generate novel context pairs outperforms these state-of-the-art GANs.