Scaling Activity Discovery and Recognition to Large, Complex Datasets
MetadataShow full item record
In the past decade, activity discovery and recognition has been studied by many researchers. However there are still many challenges to be addressed before deploying such technologies in the real world. We try to address some of those challenges in order to achieve a more scalable solution that can be used in the real world. First, we introduce a novel data mining method called the continuous Varied order Sequence Mining method (DVSM). It is able to discover activity pattern sequences, even if those patterns are disrupted or have varied step orders. We further extend DVSM into another data mining method called the Continuous varied Order Multi threshold activity discovery method (COM). COM is able to handle issues such as rare events across time and space. Furthermore, for discovering patterns in a real time manner, we extend COM as a stream mining method called StreamCOM. In addition to discovering activity patterns, we propose several methods for transferring discovered patterns from one setting to another. We propose methods for transferring activity models of one resident to another, activity models of a physical space to another, and activity models of multiple spaces to another. We also show a method for selecting the most promising sources when multiple sources are available. In order to further expedite the learning process, we also propose two novel active learning methods to construct generic active learning queries. Our generic queries are shorter and more intuitive and encompass many similar cases. We show how we can achieve a higher accuracy rate with fewer queries compared to traditional active learning methods. All of our methods have been tested on real data collected from CASAS smart apartments. In several cases, we also tested our algorithms on various other datasets.