Human Mobility Pattern Learning And Trend Prediction

With the boom in mobile devices and the Internet in recent years and the prevalence of location methods and technologies,vast amounts of mobility data have been collected,facilitating the study of human mobility patterns.The study of human mobility has important implications for traffic management,urban planning,and epidemic modeling.A large number of researchers have attempted to uncover the movement patterns embedded in human mobility data and further predict their movement trends.Also,these studies have contributed to the development of location-based services and related applications.This thesis focus on two specific tasks,personalized travel recommendation and urban traffic prediction,to mine the embedded mobility patterns and predict their future trends by deep learning methods.Personalized travel recommendation generates a sequence of places of interest for a specific tourist based on user-specific constraints.Existing personalized travel recommendation methods are either based on heuristics,simulate travel processes with the help of stochastic processes,or use recurrent neural networks to learn tourists’ visit preferences and transition patterns.These approaches rely on historical travel trajectories to train the model and use the next visited place of interest as a supervisory signal.However,due to the consistent data sparsity problem of travel trajectory data and the long-tailed distribution of places of interest according to their popularity,existing methods cannot fully capture the consistency of preferences and thus cannot effectively recommend customized trips to different users.In this thesis,a self-supervised approach is used to extract effective knowledge and supervised signals from travel trajectories,and propose a CTLTR model based on contrast learning.The CTLTR model exploits the inherent dependencies among places of interest and travel intentions embedded in travel trajectories to discover additional knowledge,and augments sparse trajectory data with a data enhancement method that can effectively mine transfer intentions.The CTLTR model provides a generic paradigm to describe the data correlations inherent in trajectories,while addressing the problem of difficult to detect implicit feedback and weak supervisory signals by learning a robust representation applicable to travel.The model introduces a hierarchical iterative encoder-decoder to identify tourists’ intentions and discover sub-trajectory semantics and their sequence patterns by maximizing mutual information.Urban flow prediction has an important role in traffic scheduling,urban management,and public safety.Previous approaches to urban flow prediction use convolutional neural networks to learn spatial correlation and recurrent neural networks to capture temporal dynamics in urban flow.Recent research work using graph neural networks to analyze urban traffic and predict flows has made promising progress.However,these methods are too simple and straightforward in constructing graphs,do not take into account implicit semantics in traffic data,and model the connections between graph nodes without considering important temporal attributes,thus failing to effectively analyze the complex spatiotemporal interactions in urban traffic.In this thesis,a deep learning model–Diff UFP is designed for predicting urban traffic based on graph neural networks.The model simulates the movement of urban crowds by constructing a spatio-temporal graph.A dynamic node representation has been designed to fully capture the temporal dynamics.the Diff UFP model uses a conditional denoising diffusion-based approach to generate the adjacency matrix of the graph from an implicit semantic space,enabling the model to fully exploit external factors to capture the shifts of flows in time and space.In this thesis,full experimental validation is conducted on several large publicly available datasets.By comparing with several common benchmark models,it is demonstrated that both models proposed in this thesis are effective in learning human movement patterns and predicting movement trends.