Spatio-temporal Feature Enhanced Semi-supervised Adaptation for Activity Recognition in IoT-Based Context-Aware Smart Homes

One of the most important applications for Internet of Things (IoT) is smart homes, where user activities generate lots of data from the interaction with IoT-based home environments. These rich interaction data can be further processed and modeled as useful contexts thus for smart homes to provide appropriate services accordingly, which in turn makes activity recognition (AR) an essential part. Traditionally, smart homes will train an AR model for all the user behaviors in advance, and then use semi-supervised learning to update AR models to deal with the changes of user behaviors while minimizing user efforts. However, if a user behavior changes too much for its original AR model, semi-supervised learning may fail to select representatives from its resulting activity instances for model adaptation, and thus causing poor performance of smart homes. In this work, we propose an approach to make use of spatial features together with temporal features to further discover more useful representative activity instances for semi-supervised learning to do AR model adaptation. Our experimental results demonstrate the effectiveness of the proposed approach.