| Abstract |
In recent years, mobility prediction has attracted much attention. Prediction methods include two steps, extracting spatial-temporal features to convert the trajectory to location sequences and constructing a model to make further predictions. Traditional methods often define the location as grids or points of interest (POIs) in mining spatial-temporal features. But these methods may not perform well in prediction because of losing detailed information of trajectories. Thus, a novel location semantics is necessary to compress detailed trajectories. In this paper, a path semantics extraction method, PathExtractor was proposed to extract typical paths and build path sequences, which contains complete information of trajectories. Furthermore, to verify that path sequences can effectively express movement patterns, the prediction is performed by constructing a recurrent neural network model. Finally, in order to evaluate the application value of path semantics, path similarity is used as performance indicator, and experiments prove the accuracy of path prediction and geographical precision higher than others. © 2020 IEEE. |
