Next-location Prediction In Human Trajectories Based On Trajectory Intention Augmentation

As one of the most valuable research hotspots in data mining,human mobility prediction is an essential task for various location-based services such as personalized PoI(Point-of-Interest)recommender systems,location-sensitive advertisement,and traffic management.With the development of mobile internet,popularization of sensor devices,and the growing ability of data collection,the location prediction problem becomes feasible and crucial.However,predicting human trajectory is not trivial because of two challenges: 1)the heterogeneity and sparsity of mobility data;2)mobility transits regularities manifest closely time-dependent and high order complexity.Most existing approaches produce a prediction by mining behavior patterns or constructing a sequential-model,which performs fine in high precision movement data like GPS location trajectory.However,in real-life conditions like human mobility trajectory,the performance of these approaches is not promising due to the above-mentioned problems of mobility data.We propose a deep learning model AI-RNN(Augmented-Intention Recurrent Neural Network)to predict the next location of the human trajectory by augmenting trajectory intention.To augment intention,we present three strategies of mobility topography construction,including random selector,direction-oriented selector,and probability selector.Then we leverage graph convolution network to get augmented location embedding under graph view,which effectively encoding the mobility pattern to augment the recurrent network for mobility prediction.Experimental results on two typical real-life datasets justify the effectiveness of the proposed AI-RNN model,our model achieves the best performance in three different datasets for top-k next position prediction.Furthermore,this thesis also analyzes other details in experiments such as trajectory regularity,collection time,and trajectory length and presents the direction of future improvements.