Research On Methods Of Trajectory Data Mining

With the popularization of mobile communication devices and the rapid development of GPS positioning technology,more and more means are available to acquire trajectory data generated by users and vehicles during the mobile process.The value behind the massive trajectory data has caused extensive research in the academic community and has also yielded a lot of research results.In particular,with the continuous emergence of mobile phone application software based on geographic location services,it is very important to accurately and reliably mine knowledge patterns in trajectory data to provide users with a better experience.The paper has two aspects of trajectory data mining work.First,we propose a user teleco trajectory recovery algorithm based on the LDA(Latent Dirichlet Allocation)topic model.Second,we propose two methods for predicting the destination of a sparse trajectory.The main contribution of our work is listed as follows:(1)For the current trajectory missing recovery algorithms,most of them need to introduce road network data,and it is difficult to obtain road network data.We proposed a telco trajectory recovery algorithm based on the LDA topic model.By preprocessing the user's telco trajectory in time and space,mapping generated a spatiotemporal word matrix.Then we divided the train data and test data into this matrix to transform the problem of telco trajectory recovery into matrix missing.The experimental results show that the accuracy of telco trajectory recovery using LDA topic model is increased 17% over F1-score@5 compared to the traditional matrix completion algorithm NMF(Non-negative Matrix Factorization),and the training time is reduced by 37%.(2)For most traditional methods predict the destination of a given trajectory by calculating the similarity between two trajectories,this algorithm does not fully consider the drawbacks of backward and forward linkages between trajectory time series,leading to a larger error in the prediction result.Therefore,we proposed a sparse trajectory destination prediction algorithm based on Markov Model.Meanwhile we investigated a grid division method based on K-d tree for the sample space of moving object motion.The experimental results show that compared with the trajectory similarity algorithm,the accuracy of the trajectory destination prediction algorithm based on the Markov model is increased by 46%.(3)For the traditional LSTM model predicting trajectory destination algorithm,ignoring the prior knowledge of trajectory destination distribution and using only the trajectory of a single direction change information,the training process is difficult and cannot fully utilize the trajectory data context information,leading to low prediction accuracy.We improve two existing defects in the trajectory destination prediction algorithm of the traditional LSTM model.First,we use the Mean Shift clustering algorithm to cluster all the destination of the training trajectory,divide the city into several areas and the center point of the corresponding area,and then add the Softmax layer behind the traditional LSTM model,so that we can predict the probability of belonging to each area,which is obtained by the probability and the center of the corresponding area.Second,we replace the unidirectional LSTM in the traditional LSTM model with a bidirectional LSTM so that context information in the trajectory data can be learned more effectively.The experimental results show that,compared with the Markov model and the traditional LSTM model,which only consider the state of the last moment,the accuracy of improved LSTM model for predicting the destination of the trajectory is increased by 13% and 28%,respectively.