Exploring semi-supervised and active learning for activity recognition

In recent years research on human activity recognition using wearable sensors has enabled to achieve impressive results on real-world data. However, the most successful activity recognition algorithms require substantial amounts of labeled training data. The generation of this data is not only tedious and error prone but also limits the applicability and scalability of today's approaches. This paper explores and systematically analyzes two different techniques to significantly reduce the required amount of labeled training data. The first technique is based on semi-supervised learning and uses self-training and co-training. The second technique is inspired by active learning. In this approach the system actively asks which data the user should label. With both techniques, the required amount of training data can be reduced significantly while obtaining similar and sometimes even better performance than standard supervised techniques. The experiments are conducted using one of the largest and richest currently available datasets.