Hierarchical destination prediction based on GPS history

Understanding and predicting destination of a trip is a crucial component of location based services. Traditional destination prediction work mostly focus on mining mobility patterns from frequently been locations. However, location transition patterns are not regular enough to provide favorable predicting results. Meanwhile, it could only be used when a user has enough movements in a location. In this paper, we propose a hierarchical model which predict what to do first and where to go in next. We first demonstrate that activity transitions are more regular than location transitions. Then we employ a Hidden Markov Model (HMM) based predicting approach which takes user's activity transition into account. We introduce a supervised way to learn parameters for HMM. Experimental results show that hierarchical prediction scheme could improve accuracy of pre-destination. Hierarchical model could perform well in some situations that traditional methods are of poor accuracy.