A Study Of Location Prediction Based On Activity Pattern

With the development of mobile Internet, user's location plays an increasingly import role in many kinds of mobile services. Beyond the current location, predicting the individual's next location can enable many novel services or applications, such as targeted advertising, location-based social networks and smart navigation. Although extensive studies about location prediction have been carried out, the traditional models make location prediction by only exploiting the individual's transferring pattern between the visited locations. The models could encounter “cold start” problem when the individual's trajectory is sparse, and perform erratically when the individual changes his activity area. Therefore, for robust location prediction, it is necessary to improve the traditional models.Contrasted with the traditional models, we aim to exploit the activity pattern for location prediction. First, we extract the temporal-spatial features associated with the user's activity from the current context, and apply the classification model to identify the user's current activity. We then model the common activity pattern and the individual's activity pattern, respectively, and combine them for predicting the individual's next activity. Finally, on the basis of the inferred next activity, Bayesian inference model is applied to predict the individual's next location.Using the trajectory dataset released by Microsoft Research Asia, we compare the advanced model with the traditional Markov models. The experimental results demonstrate that the advanced model can still perform well when the individual's trajectory is sparse, and can realize the smooth upgrade of prediction performance during the process of trajectory accumulation. Furthermore, by analyzing the influence to the model of the common activity pattern and the individual's activity pattern, respectively, it's found that the common activity pattern can help the model effectively avoid the “cold start” problem, and the individual's activity pattern can ensure high prediction accuracy for the model when the trajectory is dense.