Representation Learning-Based Trajectory Prediction

The rapid development of Location-based Social Network(LBSN)has penetrated into every aspect of our lives.Understanding human trajectory patterns is the basis of research such as recommendation systems.In recent years,Recurrent Neural Network(RNN)has achieved success in dealing with time series problems and it has also achieved significant results in predicting user trajectory behavior.User trajectory behavior prediction is based on mining user trajectory actions and predicting.The current problems of user trajectory behavior prediction are the sparseness and heterogeneity of data;the complex time dependence of user trajectory behavior;the periodicity of user trajectory behavior and so on.This thesis mainly studies trajectory prediction algorithms based on representation learning,which includes trajectory owner prediction and user trajectory next location prediction.In terms of trajectory owner prediction,previous related studies did not have better fusion of location node information and were unable to extract relatively important features for trajectory prediction;in terms of the next location prediction of the user trajectory,the location node vector without emphasis represent learning research and prediction among all exiting nodes.Although some results have been achieved,the effect is still not very good.Through researching the characteristics of different prediction problems,designing suitable representation learning methods and performing lowdimensional feature representation of location nodes and trajectories;Design a suitable neural network model to extract trajectory features.The main research contents of this thesis are as follows:(1)A trajectory owner prediction algorithm based on Convolutional Neural Network is proposed.Construct a relationship network graph of the location nodes in the trajectory,obtain the trajectory sequence by biased random walk,and then generate a lowdimensional real-valued vector representation of the location nodes;considering that the trajectory owner classification lies in the difference between different trajectories,therefore,the feature extraction of the trajectory through the Convolutional Neural Network;finally,realize the owner classification of the trajectory.(2)A next location prediction algorithm based on trajectory and user representation learning is proposed.The low-dimensional real-value vector representation of the location node is obtained by fusing the neighbor nodes feature information of each location node.After learning the feature representation of the trajectory,consider constraining the next location prediction range from all location nodes to the neighbor nodes at the end of the trajectory to improve the accuracy of the next location prediction.Generally speaking,this thesis mainly includes two aspects: representation learning of location nodes in the trajectory and feature extraction of the trajectory.At the same time,the trajectory preprocessing and the trajectory next location prediction constraints are studied in the next location prediction of the user trajectory.Through a large number of experimental comparisons on real data sets,it is verified that the method proposed in this thesis has greatly improved the accuracy of trajectory owner prediction and user trajectory next location prediction.