Deep Learning-based Location Prediction Algorithm Research And Application

In the era of big data,with the rapid development of Location-Based Services(LBS),we can obtain large-scale geographical location information of mobile objects from intelligent devices equipped with GPS devices.Thus,location prediction on big data has gained more and more attention by researchers at home and abroad.However,the existing location prediction methods faces three research challenges: firstly,most of them are generally based on continuous data such as Global Positioning System(GPS)data to analyze trajectory rules,and lack of effective analysis methods for discrete,missing semantic information or incomplete trajectory data.Secondly,the existing historical trajectory mining methods based on single-user,have not yet been fully considering the important influence of the relevance between objects for location and led to the cold start problem which has not been solved effectively.Thirdly,contextual factors(such as weather and traffic conditions)have not been effectively considered for the influence of location prediction,thus they are unable to meet the diversified needs of users for location prediction.In order to solve these challenges,this thesis proposes a Deep Learning approach for Next Location Prediction(DLNLP)to solve the problem of vehicle location prediction under complex and variable environment of the city.The main contents of this method include: Firstly,we process the vehicle passing records and generate the vehicle trajectory,then convert the trajectory into the discrete location sequence of the vehicle as the input of the algorithm.Secondly,analyze the rules of moving vehicle behavior,and consider the changing trend of the movement behavior under the contextual factors,then obtain the multidimensional features of vehicle trajectory.Thirdly,construct the location prediction method based on deep learning and use the advantages of combining Convolutional Neural Network with Deep Bidirectional LSTM to fully learn the local and global information of vehicle trajectory.Finally,the validity of the DLNLP algorithm was experimented on 197,021,276 vehicle-passing records of the Xiamen city in March 2017.Experimental results show that compared with the existing methods,DLNLP algorithm has higher prediction accuracy.The main contributions of the thesis include: firstly,we propose the methodology of vehicle location prediction and obtain the multidimensional features of vehicle trajectory.Secondly,we effectively consider the changing trend of the movement behavior under the contextual factors,then we can research the dynamic change rule of vehicle trajectory.Thirdly,we propose the location prediction algorithm based on deep learning and establish feature learning model to fully learn the behavior rules of moving objects.The method proposed in this thesis not only provides a new idea for vehicle location prediction,but also facilitates the traffic management department to provide big data decision support,such as carrying out traffic management and alleviating the congestion.