Cost sensitive GPS-based activity recognition

GPS-based activity recognition is extremely important for high-level analysis and location based services. Trajectories of people are highly imbalanced from spatial and temporal perspectives. Many existing researches achieve good results on recognizing activities with lots of GPS logs, such as working and staying at home. However, these approaches usually fail at activities with few trajectory records. In this paper, we propose a cost sensitive GPS-based activity recognition model to improve accuracy of minority activities which could imply users' personal preferences. The approach aims at providing more balanced results. We first propose a cost function to measure spatial and temporal regularities of each activity on a stay point. Then we incorporate cost function into activity recognition algorithm. We take hidden Markov model as an example in this study. Experiments show good performance of our approach in several evaluation metrics. It could provide more balanced and valuable activity recognition results from GPS trajectories.