Detecting novel class for sensor-based activity recognition using reject rule

Abstract

Many researches have shown that data streams are continuous and changeable which make them hard to be classified accurately. The major difficulty in data classification is concept evolution, namely, novel class detection. Learn++ group methods are normally employed for stream data, however, these methods hardly handle the novel class detection problem. Therefore, in this paper, we introduce an approach that combines Learn++.NSE with reject option and our research findings can be used in activity recognition whose data streams are collected by body-worn sensors. In our experiment, the proposed approach shows better performance than Learn++.NSE algorithm.