Location Prediction Methods Based On Dynamic Relationships

Trajectory data record the mobile information of the moving objects in temporal-spatial space.In recent years,with the rapid growth of positioning technologies and mobile devices,various kinds of moving objects can be traced all over the world,which greatly promotes the development of the location-based services.Current location-based applications react to the user's current location.The progression from location-awareness to location-prediction can enable the next generation of proactive,context-predicting applications.This dissertation takes moving objects' trajectory data as research object and considers the prediction of moving objects' next location as main goal.The main research works are listed as follows:(1)Most of the existing personal location prediction methods pay more attention to build the activity models that reflect the mobile preference among the locations of interest,negelecing the similarity of the routes between these locations.The dynamic moving data could help to build a better movement model under finer spatial granularity.This method not only introduces activity model to store transition information between the adjacent activities,but also predict the destination according to the real-time moving path.Finally,two types of fusion stratregies are used to combine the prediction results based on historical activies and route information.(2)The personal location prediction methods may suffer data sparity problem and the mobile bahaviours are influenced by the movement of his friends.Hence,in order to solve the problems about the inaccurate of the similarity measurement between users and the dynamic changes of the social strengths,this paper presents a location prediction method based on dynamic social ties.The time is divided by the absolute time to mine the long-term changing trend of users' social ties,and then the movements of each week are projected to the workdays and weekends to find the changes of the social circle in different time slices.The segmented friends' movements are compared to the history of the user with cross-sample entropy to discover the individuals who have the relatively high similarity with the user in different time intervals.Finally,the user's historical movement data and his friends' movements at different times which are assigned with the similarity weights are combined to build the hybrid Markov model.(3)The mobile behaviours of the moving object is a complex process,which needs an enormous amount of motion hypothesis to capture the movement patterns.Moveover,with the difference of the trajectory data type,mobile preference of the moving object and the application,we should select various movement model.However,it is hard to select a suitable model under different environment.This paper presents a location prediction method based on multilayer evaluation.Firstly,we design a large number of the base predictors.According to the prediction accuracy and the diversity between the prediction results in the trainning set,we adaptively select a set of predictors.Next,an evaluation model is built for each selected predictor,which is used to estimate the probability that the predictor could obtain the true result.Finally,A weighted voting method is used to combine the results of the selected predictors and get the final result,where the weight is the probability computed by the evaluation model.