Context-based Markov Model toward Spatio-Temporal Prediction with Realistic Dataset

This paper describes a method that simultaneously predicts the next visiting location and time of mobile users, i.e., spatio-temporal prediction (STP) from global positioning system (GPS) log dataset acquired from users' smartphones. The GPS dataset used in previous researches of STP have two characteristics, dense data and periodically collected. However, they are unrealistic assumptions since industrial provided application requires minimum number of GPS data from users in order to analyze users' context. The data sparsity and sampling rate acquired from users' daily lives with mobile devices are totally different with each user. So the data accumulated in server is not dense and not periodic. We called this problem spatio-temporal prediction with realistic dataset (STP-RD) in order to identify from original STP. To solve STP-RD, the Context-based Markov model (CbMM) which considers the convergence of heuristics in individual user's contextual information is proposed. Based on the Markov model approaches, CbMM has the unique idea which takes into consideration detailed features such as the time when data were acquired and the day of the week. For variations in data sparsity and the sampling rate which depends on individual users, CbMM detects the optimal converging state depending on the prediction. A state-of-the-art algorithm was compared by using smartphone users' location data accumulated through a mobile service of Yahoo Japan Corporation. The results indicated CbMM achieved the significantly higher accuracy than comparative approaches.