User Location Prediction Based On Incomplete Trajectory Sequences

The wide application of mobile devices and the arrival of 4G and 5G era greatly facilitate People’s Daily life,but also bring a lot of data,the user’s trajectory sequence is one of the most important kind of data.User location prediction is based on the user’s trajectory sequence,analyzes the user’s intention and preference,and then predicts the next possible location for the user to visit.At present,user location prediction has been widely used in various applications,such as tourism route planning,interest point recommendation,location completion prediction and so on.Traditional location prediction generally relies on a large number of sequence records.Typically,for a single user,a hundred or even hundreds of trajectory sequence records are read to analyze the user’s intention and preference,and then predict the user’s next visit place.However,such complete sequence records are not only difficult to collect,costly to compute,but also have a lot of noise.Meanwhile,under the background of privacy protection,the trajectory sequence records of users become increasingly fragmented,so the collected trajectory sequences are incomplete.How to use these incomplete trajectory sequences to accurately predict the user’s location is an urgent problem to be solved.Under the above background,this paper intends to predict the user’s location by the following methods:First,only incomplete trajectory sequences with a length of several to a dozen are used for prediction.Such data is not only easy to collect,but also requires less calculation and avoids privacy leakage.At the same time,it can also accurately predict the user’s location.In order to solve the problem of insufficient information of incomplete trajectory sequences,this paper uses a more elaborate modeling method to fully represent a single trajectory sequence record from the perspective of location itself and user’s intention and personalized preference.Second,the change of time characteristics of trajectory sequence records is considered.In fact,the functional of each location varies with different visit time,and the personal value of this location to the user also changes.Aiming at the problem that the existing models do not consider enough the modeling of visit time.In this paper,the visit time is represented,joint location representation,together generate a single record representation,enrich the record representation semantics,and improve the prediction effect.Third,a spatial-temporal factor called stay-time is used to express the different meanings of different locations for users in the more intuitive way,so as to reflect the personalized intention and preferences of users.Stay-time is an intuitive sign reflecting the user’s intention and preference.In a trajectory sequence,whether complete or incomplete,the longer the user stays in a certain location,the more important the location is to the user,and the closer the relationship between this location and other locations is.Therefore,stay-time can help to explore the personalized intention and preferences and to discover the connections between trajectory sequence records,laying a better foundation for further prediction.Based on the above discussion,this paper proposes a hybrid Spatio-Temporal Preference Network for incomplete trajectory sequences to predict user locations.Then,a large number of experiments are carried out on two public data sets and one collected data set: the comparison experiment with the existing model proves the superiority of the proposed model;the ablation experiment proves the role of each part of the proposed model;the hyperparameter experiment shows how various parameters affected the performance of the proposed model;the interpretability experiment demonstrates that the proposed model has certain interpretability through visualization method.