Personalized Activity Recognition using Partially Available Target Data

Recent years have witnessed growing research on autonomous activity recognition models for use in new settings. However, it lacks comprehensive frameworks for transfer learning, specifically, the ability to deal with partially available data in new settings. To address these, we propose OptiMapper, a novel uninformed cross-subject transfer learning framework. OptiMapper is a combinatorial optimization framework that extracts abstract knowledge across subjects and utilizes it for developing a personalized activity recognition model in new subjects. A novel community-detection-based clustering of unlabeled data is proposed to construct a network of unannotated observations. The clusters are then mapped onto source clusters using a complete bipartite graph model and mapped labels are conditionally fused using a base learner to create a personalized training dataset for target user. We present two instantiations of OptiMapper. The first one is applicable for domains with identical activity labels, and performs a one-to-one bipartite mapping between clusters of source and target. The second instantiation performs optimal many-to-one mapping between source and target clusters. It finds an optimal mapping even when the target dataset does not contain sufficient instances of all activity classes. We formulate this cross-domain mapping as a transportation problem and evaluate our techniques on several datasets.