Efficient Learning-based Trajectory Generation for Predicting Future Road Network
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Trajectory generation is a key application in spatio-temporal data mining, especially in the era of generative artificial intelligence. By leveraging generative models, it becomes possible to extract temporal–spatial patterns from historical trajectory data and predict future road network behaviors. Trajectory simulation now plays an essential role in navigation, traffic optimization, and mobility analysis. From a modeling standpoint, trajectory data can be viewed as multidimensional time series integrating temporal dynamics and spatial coordinates. Through reviewing existing time series generation frameworks, we identified their limitations and developed improved approaches to advance trajectory synthesis.
Chapter 1 introduces the motivation and background of deep learning–based trajectory generation. Although existing time series generation and prediction models offer practical value, they lack unified taxonomies and evaluation standards. After examining diverse deep learning models, we found that each technique has unique advantages and limitations. Selecting an appropriate model according to data characteristics is crucial. This chapter also outlines the thesis structure and summarizes our contributions.
Chapter 2 provides an overview of generative time series models. We propose a classification system based on fundamental deep learning frameworks, categorizing generative models into four types: Neural ODE–based models, RNN-based models, VAE-based models, and GAN-based models. We further analyze the strengths and applicability of each category for trajectory simulation tasks.
Chapter 3 focuses on generative models for anomaly detection. After introducing the background and application scenarios of time series anomaly detection, we review related work in spoofing detection and dynamic graph modeling for financial transactions. We then present our method combining generative time series modeling with dynamic graph representation learning. Key components include generative dynamic encoding, pseudo-labeled graph construction, heterogeneous graph attention mechanisms, and optimization objectives. Experimental settings and ablation studies demonstrate the effectiveness of the proposed framework.
Chapter 4 introduces an efficient trajectory simulation model. After summarizing related work on trajectory generation and large language models, we present our hybrid approach that integrates a graph neural network encoder with a large language model. We describe its architecture, prompt design, and inference process, along with experimental setup and comparative evaluations.
Finally, Chapter 5 summarizes the thesis and discusses contributions to trajectory generation and road network prediction. This work systematically categorizes generative time series models, applies them to anomaly detection with graph neural networks, and proposes a novel hybrid GNN–LLM trajectory generator with strong potential for smart-city applications.
Please use this identifier to cite or link to this item: