Activity Recognition Using a Few Label Samples

Sensor-based human activity recognition aims to automatically identify human activities from a series of sensor observations, which is a crucial task for supporting wide range applications. Typically, given sufficient training examples for all activities (or activity classes), supervised learning techniques have been applied to build a classification model using sufficient training samples for differentiating various activities. However, it is often impractical to manually label large amounts of training data for each individual activities. As such, semi-supervised learning techniques sound promising alternatives as they have been designed to utilize a small training set L, enhanced by a large unlabeled set U. However, we observe that directly applying semi-supervised learning techniques may not produce accurate classification. In this paper, we have designed a novel dynamic temporal extension technique to extend L into a bigger training set, and then build a final semi-supervised learning model for more accurate classification. Extensive experiments demonstrate that our proposed technique outperforms existing 7 state-of-the-art supervised learning and semi-supervised learning techniques.