Self-taught learning for activity spotting in on-body motion sensor data

Activity spotting has shown to be a highly beneficial approach in context recognition, however lacking robustness limits its wide spread use. This work introduces the concept of self-taught learning to activity spotting, which is inspired by natural human learning. The self-taught learning concept was adapted for activity spotting, in particular, to make use of unlabeled data, which does not need to include relevant pattern events. Thus, the approach can utilise background data (NULL class), for which a large amounts of data often exist. A performance comparison of self-taught and conventional activity spotters showed the potential of this new learning approach. Furthermore, an analysis using reduced amounts of supervised training instances yielded up to similar to 15% larger performance for the self-taught spotter compared to the conventional one.