궤적 기반 모델 종합 튜토리얼 스페이스 타임 일반 지능을 향한 도전과 전망

📝 Abstract

Foundation models (FMs) have emerged as a powerful paradigm, enabling a diverse range of data analytics and knowledge discovery tasks across scientific fields. Inspired by the success of FMs, particularly large language models, researchers have recently begun to explore spatio-temporal foundation models (STFMs) to improve adaptability and generalization across a wide spectrum of spatio-temporal (ST) tasks. Despite rapid progress, a systematic investigation of trajectory foundation models (TFMs), a crucial subclass of STFMs, is largely lacking. This tutorial addresses this gap by offering a comprehensive overview of recent advances in TFMs, including a taxonomy of existing methodologies and a critical analysis of their strengths and limitations. In addition, the tutorial highlights open challenges and outlines promising research directions to advance spatio-temporal general intelligence through the development of robust, responsible, and transferable TFMs.

💡 Analysis

Foundation models (FMs) have emerged as a powerful paradigm, enabling a diverse range of data analytics and knowledge discovery tasks across scientific fields. Inspired by the success of FMs, particularly large language models, researchers have recently begun to explore spatio-temporal foundation models (STFMs) to improve adaptability and generalization across a wide spectrum of spatio-temporal (ST) tasks. Despite rapid progress, a systematic investigation of trajectory foundation models (TFMs), a crucial subclass of STFMs, is largely lacking. This tutorial addresses this gap by offering a comprehensive overview of recent advances in TFMs, including a taxonomy of existing methodologies and a critical analysis of their strengths and limitations. In addition, the tutorial highlights open challenges and outlines promising research directions to advance spatio-temporal general intelligence through the development of robust, responsible, and transferable TFMs.

📄 Content

Spatio-Temporal Trajectory Foundation Model: Recent Advances and Future Directions Sean Bin Yang1, Ying Sun2, Yunyao Cheng1, Yan Lin1, Kristian Torp1, Jilin Hu3 1Aalborg University, 2 Chongqing University of Posts and Telecomunications, 3East China Normal University {seany,yunyaoc,lyan,torp}@cs.aau.dk,sunying@cqupt.edu.cn,jlhu@dase.ecnu.edu.cn Abstract Foundation models (FMs) have emerged as a powerful paradigm, enabling a diverse range of data analytics and knowledge dis- covery tasks across scientific fields. Inspired by the success of FMs—particularly large language models—researchers have recently begun to explore spatio-temporal foundation models (STFMs) to improve adaptability and generalization across a wide spectrum of spatio-temporal (ST) tasks. Despite rapid progress, a systematic investigation of trajectory foundation models (TFMs), a crucial subclass of STFMs, is largely lacking. This tutorial addresses this gap by offering a comprehensive overview of recent advances in TFMs, including a taxonomy of existing methodologies and a crit- ical analysis of their strengths and limitations. In addition, the tutorial highlights open challenges and outlines promising research directions to advance spatio-temporal general intelligence through the development of robust, responsible, and transferable TFMs. ACM Reference Format: Sean Bin Yang1, Ying Sun2, Yunyao Cheng1, Yan Lin1, Kristian Torp1, Jilin Hu3. 2025. Spatio-Temporal Trajectory Foundation Model: Recent Advances and Future Directions. In . ACM, New York, NY, USA, 5 pages. https://doi. org/10.1145/nnnnnnn.nnnnnnn 1 Introduction Self-supervised learning (SSL), such as contrastive learning [32, 34, 36, 37], and generative learning [2, 20, 22, 39, 45], has exhibited strong theoretical foundations and remarkable empirical perfor- mance across a wide range of domains, including natural language processing [4], computer vision [28], and graph data [42]. These SSL approaches enable models to learn high-quality representations from large volumes of unlabeled data, thereby reducing reliance on costly manual annotations and improving generalization in down- stream tasks. As shown in Figure 1, trajectory foundation models (TFMs)—also referred to in earlier literature as trajectory representation learn- ing—serve as a fundamental enabler for a variety of intelligent trans- portation applications, including travel time estimation [2, 12, 21, 31, 32, 34, 36, 37, 39, 45], traffic analysis [7–9, 11, 25, 38, 40, 43], tra- jectory similarity computation [1, 3, 6, 10, 13, 19, 25, 26, 30, 44, 46], trajectory recovery [18, 30], trajectory clustering [5, 27, 29, 41], Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. Conference’17, Washington, DC, USA © 2025 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-xxxx-xxxx-x/YYYY/MM https://doi.org/10.1145/nnnnnnn.nnnnnnn GPS Trajectory Grid Trajectory Textual Trajectory Image Trajectory Trajectory Data Modality Single Modality Methods Multi-Modality Methods Data Modality Oriented Methods Contrastive Learning Generative Learning Generative Contrastive Learning Causal Learning Representation Learning Methods Trajectory Foundation Models (TFMs) Data Foundation Travel Time Estimation Trajectory Similarity Computation Trajectory Clustering Trajectory Generation Downstream Applications Traffic Analysis Trajectory Revovery Trajectory Classification Figure 1: The framework of TFMs for ST Trajectory data. trajectory classification [15, 24, 33, 37, 44, 45], and trajectory gener- ation [30, 47, 48]. In particular, a trajectory can be formally defined as𝑇= ⟨(𝑥1,𝑦1,𝑡1), (𝑥2,𝑦2,𝑡2), · · · , (𝑥𝑛,𝑦𝑛,𝑡𝑛)⟩, where each element represents a spatio-temporal point consisting of spatial coordi- nates (𝑥𝑖,𝑦𝑖) and a corresponding timestamp 𝑡𝑖. Such trajectories inherently capture both spatial movement patterns and temporal dynamics, making them a rich yet challenging modality for effective representation learning. Tutorial Overview. With the rapid proliferation of spatio-temporal data, trajectory foundation models have emerged as a critical re- search direction, attracting increasing attention from both academia and industry. This tutorial provides a systematic and comprehen- sive overview of this evolving field. We begin by formalizing the core concepts of trajectory data mining and trajectory foundation models, establishing a unified basis for subsequent discussions. The objectives

This content is AI-processed based on ArXiv data.