Incorporating unsupervised learning in activity recognition

The feasibility of applying subspace clustering--a specific type of unsupervised learning-- to high-dimensional, heterogeneous sensory input is analyzed and the correspondence between clustering output and classification input is presented.

Abstract

Users are constantly involved in a multitude of activities in ever-changing context. Analyzing activities in context-rich environments has become a great challenge in context-awareness research. Traditional methods for activity recogni-tion, such as classification, cannot cope with the variety and dynamicity of context and activities. In this paper, we pro-pose an activity recognition approach that incorporates unsu-pervised learning. We analyze the feasibility of applying sub-space clustering—a specific type of unsupervised learning— to high-dimensional, heterogeneous sensory input. Then we present the correspondence between clustering output and classification input. This approach has the potential to dis-cover implicit, evolving activities, and can provide valuable assistance to traditional classification based methods.