Stream Sequence Mining for Human Activity Discovery

During the past decade, supervised activity recognition methods have been studied by many researchers; however, these methods still face many challenges in real-world settings. Supervised activity recognition methods assume that we are provided with labeled training examples from a set of predefined activities. Annotating and hand-labeling data is a very time-consuming and laborious task. Also, the assumption of consistent predefined activities might not hold in reality. More important, these algorithms do not take into account the streaming nature of data, or the possibility that the patterns might change over time. This chapter provides an overview of the state-of-the-art unsupervised methods for activity recognition. In particular, we describe a scalable activity discovery and recognition method for complex large real-world datasets based on sequential data mining and stream data mining methods.