Research On Individual Travel Destination Prediction Based On Deep Learning

With the development of social networks,individuals generate massive check-in data,which has become an important data source in research fields such as analysis of individual mobility regularities and destination prediction.Predicting individual travel destinations has important significance in research fields of traffic flow prediction,location recommendation,epidemic prevention and control.How to obtain highefficiency individual embedding representations and how to model individual behaviors are the key research issues in this dissertation.Aiming at the research of individual travel destination prediction research,this dissertation proposes two prediction methods.(1)Prediction method based on exponential decay and neural network model.Existing models lack attention to complex patterns of individuals.In this dissertation,multi-information such as time series,event descriptions,travel locations are combined to construct an event-aware representation learning technique to obtain low-rank vector expressions of individuals in the latent space.In order to solve the problem of insufficient mining of potential information in data by traditional methods,a learning strategy called Category-Important-Decay Learning is proposed to emphasize the impact on historical trajectory on the current movement state based on exponential decay and frequent sequence pattern mining,which can obtain individual embedding representations.Finally,individual embedding representations is fed to recurrent neural network to capture the individual temporal and spatial movement regularities based on event perception,and improve the accuracy of travel destination prediction.(2)Prediction method based on attention mechanism and neural network.In order to further improve the efficiency of the model,this dissertation also designs a multi-module embedding recurrent neural network that combines the attention mechanism and the neural network.The network is composed of a historical attention module and a recurrent neural network module.Using only the sequence and time information of the individuals,the individual’s periodicity transition regularities and sequence transition regularities can be learned.We divide the individual travel trajectory into two parts: the current state and the historical state.The historical state is fed into historical attention module by designing a historical attention learning mechanism based on weighted self-learning to capture the periodicity transition regularities of individuals.In addition,the current state is fed into recurrent neural network to capture the individuals’ sequence patterns.This method can simplify the network structure and improve the network efficiency,and obtain accurate prediction results.In order to verify the effectiveness of the above two prediction methods,this dissertation conducted experiments on two representative social networks check-in datasets Twitter and Foursquare.The results show that the two prediction methods proposed in this dissertation both can improve the accuracy of prediction in the research of travel destination prediction research.