Addressing the problem of activity recognition with experience sampling and weak learning

Abstract

Quantifying individual’s levels of activity through smart or proprietary devices is currently an active area of research. Current implementations use subjective methods, for instance, questionnaires or require comprehensively annotated datasets for automated classification. Each method brings its own specific drawbacks. Questionnaires cause recall bias and providing annotations for datasets is difficult and tedious. Weakly supervised methodologies provide methodologies for handling inaccurate or incomplete annotations and literature has shown their effectiveness for classifying activity data. As a key issue of activity recognition is capturing annotations, the aim of this work is to evaluate how classification performance is affected by limiting annotations and to investigate potential solutions. Experience sampling combined with the algorithms in this paper can result in a classifier accuracy of 74% with a 99.8% reduction in annotations, with increased compute overheads. This paper shows that experience sampling combined with a method of populating labels to unlabeled feature vectors can be a viable solution to the annotation problem.