Research On Location Representation Learning Based On Spatio-Temporal Trajectory

With the imperative demand of constructing smart cities and the continuous accumulation of user spatio-temporal trajectory data,the location based urban traffic data mining has been extensively studied.People hope to have a more accurate understanding of traffic problems and demands in urban development in order to take effective traffic strategies.In addition,location-based data mining can assist with location-related business activities.In the research of location-based data mining,location representation learning is a very important fundamental research task.The goal of location representation learning is to learn an embedded and latent feature vector for each location.However,most of the existing location representation methods only consider the spatial and sequential correlations of locations in trajectories,and ignore the temporal influence that at what time a user arrives at a certain location in a trajectory.Considering that the arrival time of users can explicitly reflect the locations5 characteristics,it is necessary to embed the arrival time into the location representation learning.In this dissertation,a Time-Aware Location Embedding(TALE)method is proposed to learn the distributed representations of locations from spatio-temporal trajectories.In this method,we design a novel tree structure based on the hierarchical softmax model,and incorporate the temporal information in trajectory data into the child nodes of a multi-tree.In order to verify the effectiveness of the vectors obtained by TALE,we carry out a series of analogy experiments of the embedded vectors,and the experimental results demonstrate that our TALE model achieves better analogy accuracy.Since location representation translates the features of locations as low-dimensional vectors and the latent features hidden in these vectors can help us to better perform some location-related mining tasks,we further perform two location-based prediction tasks,i.e.,user next location prediction and location traffic flows prediction.In order to perform the two prediction tasks better,we integrate the embedded vectors obtained by location representation models into the time series prediction models as input to predict users' next location and location traffic flows respectively.We collected some mobile phone signaling data in Beijing within five consecutive working days as the spatio-temporal trajectory dataset.The experimental results show that,compared with the existing baseline methods,the proposed prediction models incorporating location latent representations can yield better performance.And for different location representation learning methods,the TALE model can further significantly improve the prediction accuracy on the two tasks,which indicates that TALE can better learn the implicit features of locations.