Research On Sematntic Location Prediction Technology Using Variable Order Markov And LSTM

With the fact that smart phone sensor can be more and more convenient to collect user location information,topics about how to use location information to discover users' hobby and mobile habits are current technology hotspot.One of the research fields is predicting the user's future location based on his history trajectory.Location prediction has drawn much attention for its important applications in many fields such as take-away service,taxi service,real-time public transport system and advertisement.At present,there are many academic achievements in the field of position prediction.However,some problems occurred:the stay point extracted from the original trajectory has poor performance;it is difficult to predict the user's un-visited location;difficulties in building models with other features;the traditional method has dissatisfied performance in long position sequence;the prediction accuracy is lower than user's expectation.Therefore,this paper further studied those topics and proposed two improved solutions to solve the mentioned problems and improve the accuracy of model prediction.The paper proposed the first improvement model based on the position prediction of variable-order Markov model.The core of this model is Markov model with dictionary tree,PPM.This paper applied many methods to convert original data set to location sequence,such as noise point filtering,stay point extraction,clustering,and location conversion.In order to improve the modeling performance,this paper combined frequent trajectory tree with PPM model.Establishing frequent trajectory tree according to user's trajectory.According to user location frequency,this paper applied AP algorithm to cluster users into different groups.This paper put the position sequence of each user group into the PPM model for training.This paper brought the current user's trajectory into the model,and location with the highest probability is the result of the prediction.The paper proposed the second improvement scheme based on LSTM model.For the reason that the user's future location is affected by many factors,the location,time period,and weather conditions are transformed into feature vectors in this paper.In order to increase the diversity of training tags,this paper segmented the trajectory by sliding window.Finally,this paper trained the prepared feature data in the LSTM model,and then it used the trained model to predict user's position.The experimental results shown that accuracy of the two prediction models is improved compared with the original model.The variable-order Markov prediction model performed better in predicting locations where users have never visited.The second prediction model shown good performance in reflecting time rule and user's intention.