The Research Of Location Prediction Based On Recurrent Neural Network

The advance of location-acquisition technologies enables people to record their location histories with spatial-temporal datasets, which imply abundant information, such as people’s interest, habits, etc. Prediction next location based on the historical trajectory has always been the hot research topic. Especially with the development of location based service(LBS), the effective prediction can’t only provide people good experience, but also high practical value and broad application scenario.In this paper, we review the research progress of the location prediction technology, summarize some existed problems:ineffectively extracting stay point (a place where people spend some time for activities) from the original trajectories; prediction accuracy is not high etc. Addressing these problems, this paper presents a new prediction algorithm based on the Recurrent Neural Network, due to the limitation of storage information and unstable, we use the Long Short Term Memory model.We first propose a multistage cluster algorithm to extracting landmark (a region which contains some stay points). The algorithm includes:1.the heuristic filtering the abnormal GPS point,2.extracting the stay points based the regional coherency principle,3.using DBSCAN algorithm to detecting significance places. Trade space for time accelerates the algorithm execution efficiency.In general, the prediction next place always has a certain context relation with the previous visited places. According to the hypothesis, landmark’s index sequences feed in our two-level model, extract the feature vector and predict the next place with maximum probability.As a result, we verify the reliability of the algorithm with the Geolife dataset which contains 18670 trajectories, acquires over 49.52% and 71.25% accuracy on different region radius, beats tradition prediction model.