Urban trajectory modeling and recovery with deep learning

Abstract

The ubiquity of GPS-enabled devices has led to an unprecedented generation of massive spatial trajectory data by various moving objects, such as vehicles and humans, within urban spaces. Extracting meaningful patterns and insights from this extensive trajectory data is crucial for addressing real-world problems and enhancing daily life experiences. This thesis focuses on distilling knowledge from urban trajectory data through innovative deep learning methods, specifically in the realms of urban trajectory modeling and recovery. In the domain of trajectory modeling, two significant scenarios are explored: Road Network Based Non-geographic-intention-aware Trajectory Modeling and Road Network Based Geographic-intention-aware Trajectory Modeling. The former investigates trajectory modeling when the destination is unknown and we propose a novel Multi-task Modeling framework for Trajectories (MMTraj), which is an end-to-end multi-task learning model for trajectory modeling. Our major contributions lie in paradigm-shifting trajectory modeling by predicting trajectory destination information through a multi-task learning strategy, where a dedicated gating mechanism refines the utilization of predicted destination information for enhanced trajectory modeling. We introduce an advanced framework architecture applying Graph Neural Networks (GNN) and a dual-transformer design. This framework leverages the complete geometrical graph topology of road networks and incorporates a multi-head self-attention mechanism for robust trajectory sequence learning. Empirical validation through extensive experiments on two real-world mobility datasets demonstrates the superior performance of our framework over baseline methods, highlighting its practical efficacy in trajectory modeling. The latter scenario deals with trajectory modeling when considering further on the geographic bearing angle information and besides this, we also add one more setting to extend it to the case when the destination is known. We introduce the extended model MMTraj+. MMTraj+ not only predicts trajectory destinations but also the critical information of bearing angles and leverages this information to guide the main task of predicting the next road segment. An innovative approach incorporating the bearing angle as an azimuthal feature in deep learning neural network-based data fusion has been investigated. Diverse modules are crafted to capture relative geo-locations and azimuthal angles, offering additional insights from the destination to the current location. These modules act as complements to destination predictions, providing supplementary guidance for the primary task. Additionally, a complementary model, termed DestTraj, is developed to cater to a specific scenario of trajectory modeling with known destinations. This broader applicability expands the utility of trajectory modeling in various applications. Both MMTraj+ and DestTraj exhibit superior performance over all existing methods in real-world datasets. Additionally, the thesis delves into the critical task of Urban Human Trajectory Recovery. Recognizing the sparse nature of human mobility data from location based services, our focus is on the retrieval of missing locations using current and historical trajectory data. To this end, we introduce a novel approach called conditional Diffusion model for Movement (DiffMove). DiffMove employs self-supervised learning techniques and conditional diffusion probabilistic models to capture complex spatial-temporal dependencies within individual trajectories. This model represents the first instance of applying a conditional diffusion approach to recover missing and discretized human trajectory data. Our innovation involves the design of key modules, including Spatial Conditional Block, Target Conditional Block, and Denoising Network Block. These modules collectively enable comprehensive modeling and exploitation of spatial temporal dependencies within conditional information, facilitating the effective recovery of missing locations. DiffMove not only captures intricate spatial-temporal dependencies but also offers a principled diffusion method for handling missing trajectory data. It provides a principled approach to handle missing data at a fine-grained resolution, showcasing significant improvements over baselines in real-life mobility datasets. In summary, this thesis contributes novel deep learning frameworks and methodologies for urban trajectory modeling and recovery, addressing scenarios with trajectories of both moving vehicles and human mobility. The proposed models exhibit superior performance, opening new horizons for trajectory analysis in urban analytics and transportation planning.