Improving human activity classification through online semi-supervised learning

Abstract

Built-in sensors in most modern smartphones open multipleopportunities for novel context-aware applications. Although the HumanActivity Recognition field seized such opportunity, many challengesare yet to be addressed, such as the differences in movement by peopledoing the same activities. This paper exposes empirical research onOnline Semi-supervised Learning (OSSL), an under-explored incrementalapproach capable of adapting the classification model to the userby continuously updating it as data from the users own input signalsarrives. Ultimately, we achieved an average accuracy increase of 0.18percentage points (PP) resulting in a 82.76% accuracy model with NaiveBayes, 0.14 PP accuracy increase resulting in a 83.03% accuracy modelwith a Democratic Ensemble, and 0.08 PP accuracy increase resultingin a 84.63% accuracy model with a Confidence Ensemble. These modelscould detect 3 stationary activities, 3 active activities, and all transitionsbetween the stationary activities, totaling 12 distinct activities