Personalized Itinerary Planning Based On Deep Trajectory Mining

With the development of social networks,people are becoming increasingly happy to post and share various information about their daily life,hence the amount of data of location information is exponentially growing at the same time.Both mining location information and revealing human mobility patterns are of vital importance for many downstream applications.For example,the research in the field of Point of Interest(POI)recommendation and traffic congestion prediction has been very popular and widely conducted.However,personalized itinerary recommendation based on mining user mobility patterns is still an important and urgent challenges remain to be solved.The mobile Internet provides users not only with unprecedented convenience,but also with extremely rich contextual information for the study of personalized itinerary planning.In recent years,various methods based on sequential modeling and presentation technology have been continually emerging,such as ranking-based matrix factorization models,trajectory prediction models based on random walks and so on.But those methods have rarely focused on the research of trajectory semantics.However,discovering abstract themes related to movement and utilizing the contextual semantics of the trajectory can provide a more comprehensive understanding of the user's movement patterns.In addition,trajectory reconstruction and trajectory prediction are two sub-tasks of personalized itinerary planning.In-depth study on those two sub-tasks can help a better grasp on the user's travel preference so as to customize the itinerary for users.The main contents and innovation of this thesis are as follows:1.This thesis proposes an itinerary planning model based on variational trajectory context perception,which mainly includes four frameworks for learning user trajectory presentation:(1)recurrent encoder,(2)variable encoder,(3)variable sub-attention layer,(4)two decoders.Aimed at itinerary planning for users,this method,which is a new example of mobile pattern mining based on learning trajectory context,begins with user trajectory information and adopts the encoder-decoder structure to solve the trajectory context learning problem at the trajectory level.2.This thesis proposes an itinerary planning model based on hierarchical trajectory presentation.The model utilizes enhanced learning based on policy networks to rationally divide each trajectory to obtain a hierarchical presentation of the trajectory.As trajectory segmentation plays an important role in mobile trajectory mining,hierarchical presentation of the trajectory can mine deeper and more comprehensive trajectory semantics and obtain more abundant mobile context information,and then can improve the performance of personalized itinerary recommendation.3.In order to verify the accuracy and effectiveness of the two methods among the two sub-tasks of trajectory reconstruction(T1)and trajectory prediction(T2),we have evaluated the methods proposed on several public datasets and compared with other existing methods,it's shown that encoding the context trajectory vector can effectively characterize the hierarchical mobility semantics and decode the implicit meaning of the trajectory as well.The experiments have proved that the proposed two models can effectively enhance the performance of personalized itinerary planning.