Periodic Activity Pattern Mining Base On Personal Trajectory

With the rapid development of location-based technologies, massive amounts of trajectory data have been collected from people’s daily life. These data contain lots of knowledge. Trajectory data mining has drawn great attention of many experts and scholars,and become one of the current research hotspot. Not only the daily activities pattern but also some important potential information can be mined from personal trajectories.Periodicity is a phenomenon which occurs frequently in our daily life. It can be used to further understand our activities and hobbies through periodic activities pattern mining. We try to make a research on mining people’s periodic activity pattern from their historical trajectory in this thesis.In order to solve the problems of uncertainty data sampling frequency and data sparseness, we propose an algorithm to detect the periodic activity pattern, which is based on probabilistic models. Because of the complexity of periodic activity, such as multi-period cross-cutting and uncertainty partial period span etc., we propose a three-stage algorithm to solve it. At the first stage, a trajectory pre-processing algorithm is proposed to extract stay points from each individual’s trajectories. At the second stage, the sequence of stay points is clustered by the density-based clustering algorithm to construct the important location set of users. And then the degree of interest for each location in the set are calculated to find out the reference spots. At the last stage, the original trajectory is transformed into binary sequence. Period of every binary sequence is detected by probabilistic measure. Then the method based on the hierarchical clustering is employed to mine the periodic activity pattern.Compared with the traditional methods on periodic activity mining, the method proposed in this thesis has two advantages. First, it can detect period automatically without additional period parameter which is determined by the users according to the actual data.Second, a probabilistic measure is adopted to detect periods. The method has thoroughly considered the uncertainties and noises in periodic activities and is provably robust to incomplete observations.The effectiveness of the approach is validated based on simulated data set and real data set collected in Geolife project. The experiment results show that the proposed methodcan mine people’s periodic activity pattern effectively and is provably robust to incomplete observations and sparse data.