MultiSource-Adversarial-Domain-Adaptation

Human activity recognition (HAR) fromwearable sensors data has become ubiquitous due to thewidespread proliferation of IoT and wearable devices.However, recognizing human activity in heterogeneousenvironments, for example, with sensors of different modelsand make, across different persons and their on-bodysensor placements introduces wide range discrepancies inthe data distributions, and therefore, leads to an increasederror margin. Transductive transfer learning techniquessuch as domain adaptation have been quite successful inmitigating the domain discrepancies between the sourceand target domain distributions without the costly targetdomain data annotations. However, little exploration hasbeen done when multiple distinct source domains arepresent, and the optimum mapping to the target domainfrom each source is not apparent. In this paper, we pro-pose a deep Multi-Source Adversarial Domain Adaptation(MSADA) framework that opportunistically helps selectthe most relevant feature representations from multiplesource domains and establish such mappings to the targetdomain by learning theperplexity scores. We showcasethat the learned mappings can actually reflect our priorknowledge on the semantic relationships between thedomains, indicating thatMSADAcan be employed as apowerful tool for exploratory activity data analysis. Weempirically demonstrate that our proposed multi-sourcedomain adaptation approach achieves 2% improvementwith OPPORTUNITY dataset (cross-person heterogeneity,4 ADLs), whereas 13% improvement on DSADS dataset(cross-position heterogeneity, 10 ADLs and sports activi-ties).