Research Of Location Prediction Based On Behavior Patterns And Public Transit

With the rapid development of the mobile Internet and the advancement of positioning technology,the next location prediction has become an important research task,which is of great significance to users and businesses.However,due to the sparsity of user check-in data and the complex correlation of related features,there are huge challenges in predicting the next location.Firstly,existing next location prediction algorithm lacks specific behavior pattern analysis for a certain location,which makes it impossible to fully model the user's historical behavior pattern and short-distance travel patterns for location prediction.Secondly,geospatial features cannot fully model the user's travel features,and existing models lack the modeling of long-distance travel features of users based on spatial geographic features.To solve the problem of insufficient modeling of users' historical behavior and short-distance travel patterns,this thesis proposes a location prediction model based on time window.Firstly,construct a check-in temporal graph via using the common timing features between the points of interest,obtain sampling sequences by random walk,learn the context information of interest points in a graph embedding method,and capture the non-personalized time sequence features between interest points.Secondly,the user's behavior pattern has the features of repeatability and reverse sequence.This thesis uses time windows to capture repetitiveness of users' historical behavior patterns,and uses the bidirectional time window to capture the user's reverse sequence behavior patterns at same time.Furthermore,user behavior patterns are highly correlated with geographic features,this thesis uses geographic feature weighting methods to model user short-distance travel patterns within the time window.Finally,the temporal dependency is further extracted through the recurrent neural network,by combining the personalized time series features and the user's historical behavior features,the next position prediction is performed.To solve the problem that geospatial features do not adequately model users' long-distance travel features,this thesis proposes a prediction model that combines the characteristics of public transportation by improving the modeling of behavior patterns within the time window.Firstly,the public transport dataset is associated with the check-in data set of points of interest to calculate the convenience of public transport between points of interest.Then,the user's travel pattern is modeled by combining the characteristics of public transportation and the geospatial characteristics.This method can capture users' long-distance travel patterns and short-distance travel patterns simultaneously,making up for the shortcoming that geospatial features can only model users' short-distance travel patterns.Experiments are carried out on a real-world point-of-interest check-in dataset.The proposed model is compared with other baselines for location prediction performance by using the evaluation indexes.Experimental results respectively prove the validity of the proposed time-window prediction model and the public transport model for the modeling of user long-distance traffic characteristics,and verifies the effectiveness of different modules in the prediction model is verified through ablation analysis.