Deep Learning And Regressive Classification Based Destination Prediction

With the popularity of GPS positioning device and Internet,it is easier to collect history data of human travelling trajectories.After collecting this kind of data,people may discover useful information in it,such as travelling routines.According to these traveling routines,we are able to predict a user's final destination by just looking at part of the user ' s travelling trajectory.If we can predict the final destination precisely,we can offer services like travelling advices and accurate advertising in commercial area,or tracking people like suspects or lost ones in city security area.We reviewed research progress on travelling destination prediction,and find following problems:similar history trajectory based prediction usually cannot predict well,as similarity between trajectories is hard to define;Markov model based prediction is better,however,Markov model assumes car travelling satisfies Markov property,which cuts apart the before-after relationship among the locations of the trajectory;models treat locations as separate particles,which also cuts apart the spatial relationship among the locations;time information is usually ignored by most models,while apparently,time is a very important factor influencing human travelling behavior.We use recurrent neural network in deep learning to predict users'travelling destinations.Recurrent neural network has been proven to be good at memorizing,which is quite suitable for dealing with tarjectory.We use classification method,by firstly preprocessing common areas of trajectories to grids of some precision,then replace trajectories with grid cells,and finally predict the destination grid cell.Embedding technology has been proven to be good at feature learning from discretely labeled data,and we will use embedding technology to learn features from grids and time.Further,we introduce the concept of regressive possibility distribution,which can denote the distance between grid cells by possibility.We use regressive possibility distribution to change plain old classification model to regressive classification model.Experiments show that,this kind of prediction can reduce prediction error significantly comparing to ordinary classification and Markov model.