Toward Foundation Models for Mobility Enriched Geospatially Embedded Objects.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Proc ACM SIGSPATIAL Int Conf Adv Inf . Author manuscript; available in PMC: 2026 Apr 13. Published in final edited form as: Proc ACM SIGSPATIAL Int Conf Adv Inf. 2025 Dec 12;2025:774–779. doi: 10.1145/3748636.3760459 Search in PMC Search in PubMed View in NLM Catalog Add to search Toward Foundation Models for Mobility Enriched Geospatially Embedded Objects Maria Despoina Siampou Maria Despoina Siampou † University of Southern California, Los Angeles, California, USA Find articles by Maria Despoina Siampou † , Shang-Ling Hsu Shang-Ling Hsu † University of Southern California, Los Angeles, California, USA Find articles by Shang-Ling Hsu † , Shushman Choudhury Shushman Choudhury ‡ Google Research, Mountain View, California, USA Find articles by Shushman Choudhury ‡ , Neha Arora Neha Arora ‡ Google Research, Mountain View, California, USA Find articles by Neha Arora ‡ , Cyrus Shahabi Cyrus Shahabi † University of Southern California, Los Angeles, California, USA Find articles by Cyrus Shahabi † Author information Article notes Copyright and License information † University of Southern California, Los Angeles, California, USA ‡ Google Research, Mountain View, California, USA ✉ Email: [email protected] Issue date 2025. This work is licensed under a Creative Commons Attribution 4.0 International License . PMC Copyright notice PMCID: PMC13070354  NIHMSID: NIHMS2157160  PMID: 41978901 The publisher's version of this article is available at Proc ACM SIGSPATIAL Int Conf Adv Inf Abstract Recent advances in large foundation models (FMs) have enabled learning general-purpose representations in natural language, vision, and audio. Yet geospatial artificial intelligence (GeoAI) still lacks widely adopted foundation models that generalize across tasks that require joint reasoning over geospatial objects and human mobility. Such tasks are crucial as mobility, along with satellite imagery, street view, and text, is a core modality for understanding the physical world. We argue that a key bottleneck is the absence of unified, general-purpose, and transferable representations for geospatially embedded objects (GEOs). Such objects include points, polylines, and polygons in geographic space, enriched with semantic context and critical for geospatial reasoning. Much current GeoAI research compares GEOs to tokens in language models, where patterns of human movement and spatiotemporal interactions yield contextual meaning similar to patterns of words in text. However, modeling GEOs introduces challenges fundamentally different from language, including spatial continuity, variable scale and resolution, temporal dynamics, and data sparsity. Moreover, privacy constraints and global variation in mobility further complicates modeling and generalization. This paper formalizes these challenges, identifies key representational gaps, and outlines research directions for building foundation models that learn behavior-informed, transferable representations of GEOs from large-scale human mobility data, as well as static contextual information such as points of interest, object shapes and spatio-temporal semantics. Keywords: foundation models, GeoAI, representation learning, spatio-temporal modeling, human mobility, spatio-temporal reasoning 1. Introduction Large foundation models have transformed natural language processing and computer vision by enabling models to learn contextual, general-purpose representations of words and images. These models, trained on vast amounts of publicly available data, can capture complex semantic and structural relationships and solve a wide range of downstream tasks with minimal task-specific supervision. The core idea behind these successes is that of a single model serving as a flexible backbone for many applications by leveraging transferable representations [ 6 ]. Despite these advances in other domains, GeoAI has yet to see comparable progress [ 12 , 15 , 24 , 31 , 46 ]. We argue that a key challenge lies in learning representations for geospatial objects that capture their differences in geometry, from points (e.g., business locations) to polylines (e.g., street segments) to polygons (e.g., building footprints), as well as their semantic attributes (e.g., building function). These objects are essential for globally effective geospatial reasoning tasks, such as determining whether a coffee shop is located within a mall or computing the distance from a point of interest (POI) to the nearest road. We introduce the term Geospatially Embedded Objects (GEOs) to refer to these entities: points, polylines and polygons situated in geographic space, enriched with semantic context, that are integral to spatial reasoning. The analogy to language is intuitive: just as words derive meaning from their context within sentences, GEOs can gain contextual meaning from patterns of human movement over time, with sequences of interactions with GEOs resembling sentences [ 9 ]. This analogy has motivated efforts to apply LLMs to learn GEO embeddings from trajectories [ 32 , 33 ]. However, the analogy breaks down in practice: language-based techniques struggle to capture the unique characteristics of GEOs. Unlike words, GEOs are embedded in continuous space, vary in scale, exhibit complex temporal dynamics, and are visited sparsely and unevenly. Additionally, modeling GEOs introduces distinct challenges, such as privacy concerns in mobility data and limited generalizability across cities due to differences in urban structure and movement patterns. These differences call for rethinking existing modeling paradigms and developing new approaches specifically designed for the geospatial domain. In this paper, we outline a vision for GEO-centric foundation models. We decompose the end-to-end modeling pipeline and analyze unique challenges for which existing language-based methods fall short. We also highlight key considerations, including privacy, transferability, and interpretability, that are essential for building robust, general-purpose representations of GEOs. 2. Unique Modeling Challenges 2.1. Mobility-to-GEO Attribution The first step in learning mobility-enhanced representations of GEOs includes converting raw GPS traces into sequences of geospatial objects, a process we call GEO attribution . This step identifies which GEOs (e.g., POIs, roads, or neighborhoods) are visited, passed by, or generally associated with a given trip. Conceptually, GEO attribution loosely parallels tokenization in NLP, where raw input is segmented into discrete tokens. However, unlike language, where tokens are well-defined and drawn from a fixed vocabulary, geospatial “tokens” must be inferred from continuous spatial traces. Mapping GPS points to GEOs is an ambiguous, dynamic, and task-specific process. GPS data is often noisy or imprecise, making it difficult to determine which GEOs are truly relevant to a trip. Yet, even with clean trajectories, attribution poses several modeling challenges. For example: Should we include only explicitly visited GEOs, or also those merely passed nearby? This decision has important implications. Including too many nearby but irrelevant GEOs could lead to over-attribution, while failing to capture brief but meaningful stops could result in under-attribution (e.g., missing a transit hub due to short dwell time). Even the attribution methods themselves vary widely. For instance, map matching align GPS points to road networks [ 34 ] while POI attribution attempts to associate visits based on spatial and temporal cues. For the latter, existing methods rely mostly on heuristics such as fixed dwell-time thresholds, nearest-neighbor assignments, and spatial buffers, which constrain accuracy [ 35 , 36 , 40 ]. The challenges increase when GEOs serve multifunctional roles (e.g., a transit hub that is also a shopping center), or when semantic importance outweighs proximity; for example, GPS points recorded near a university may be closest to coffee shops and adjacent amenities, yet a longer dwell time suggests the visit should be attributed to the campus [ 37 ]. Lastly, unlike deterministic tokenization in language, GEO attribution is context-dependent, as the same trajectory may yield different GEOs depending on whether the goal is routing or behavior analysis. All these challenges make GEO attribution an inherently uncertain and ill-defined problem, that needs to be solved to enable effective GEO representation learning. 2.2. GEO Encoding Once GEOs have been identified, the next step is to convert them into fixed-length vectors suitable for downstream learning tasks. This step, formally referred to as encoding, is loosely analogous to word embedding in NLP, where models like Word2Vec [ 30 ] capture semantic similarity based on co-occurrence in text. However, GEO encoding is significantly more complex due to the multimodal nature of geospatial data. GEOs are defined not only by their location in space, but also by their functional roles (e.g., serving as a school or a hospital) and by temporal patterns of interaction, such as when and how frequently they are visited. These diverse attributes must be jointly captured to produce representations that are both meaningful and generalizable. The remainder of this section discusses the challenges and representative methods for encoding GEOs along these three key dimensions: spatial, contextual, and temporal . 2.2.1. Spatial Dimension. Spatial characteristics are central to the identity of GEOs and must be explicitly preserved in their representations. However, encoding spatial data presents unique challenges. First, encoders must support heterogeneous geometries , including points, polylines, and polygons. Most existing approaches have focused on point geometries [ 25 ], with comparatively limited attention given to polylines and polygons [ 27 , 47 ]. Although geospatial objects can be converted into alternative formats like images or text to fit standard machine learning pipelines [ 5 , 7 , 8 , 18 , 45 ], such transformations often discard critical spatial information, such as the object’s exact position in space, which can degrade performance on downstream tasks. Recent efforts like Poly2Vec [ 39 ] represent early progress toward a unified encoding framework that preserves spatial characteristics across diverse geometry types. Second, geospatial data spans multiple spatial scales , from neighborhoods to cities and regions, requiring representations that remain robust across varying resolutions. To that extent, some methods operate on a fixed grid scale based on task assumptions (e.g., zip code level prediction) [ 1 , 48 ], while others adopt hierarchical schemes to encode information across multiple levels [ 8 , 19 ]. These approaches remain sensitive to grid design and often fail to generalize across scales. Multi-scale encoders offer more flexibility, but current designs are limited to point geometries [ 26 ]. Third, spatial encoders should capture a rich set of spatial properties . While most existing methods primarily focus on distance-based proximity [ 16 , 18 ], they should also capture topological (e.g., adjacency, containment) and directional relationships as well as structural characteristics, such as the curvature of a road or the footprint complexity of a region. These properties are essential for enabling geospatial reasoning tasks that go beyond proximity, such as identifying whether a building lies within a hazard zone, or determining whether two roads are connected. Despite their importance, such properties are rarely captured or evaluated in existing pipelines [ 39 ]. Addressing these gaps calls for encoders that unify geometry types, support continuous spatial input, and capture rich relational and structural properties. 2.2.2. Contextual Dimension. GEOs are often associated with rich contextual information that provides semantic grounding beyond geometry alone. This includes object level functional roles and categories, such as determining if a GEO is a hospital, a residential building, or a main road, as well as geometry-specific metadata, like the number of floors of a building or traffic volume. These features, often sourced from OpenStreetMap, government records, or remote sensing, are essential for understanding a GEO’s role and should be embedded directly into its representation [ 5 , 11 , 19 , 23 ]. Context also extends to the neighborhood level , where features such as the distribution of nearby POI types capture a GEO’s functional role within its broader environment. Some studies aggregate neighborhood features using fixed-radius buffers [ 19 , 42 ], spatial attention mechanisms [ 8 , 26 ] or graph-based approaches [ 13 , 44 , 49 ]. These signals are crucial for capturing urban structure, functional zoning, and patterns of human activity. Despite their importance, semantic attributes are often treated as standalone metadata, appended to GEO’s representation without modeling their interaction with the object’s geometry. This limits models ability to capture how meaning arises from the interplay between spatial features and semantics. A key challenge is to design representations that reflect this interdependence. While a few studies explore joint learning between geometry and semantics [ 5 ], approaches that explicitly model these relationships remain limited. 2.2.3. Temporal Dimension. In LLMs, sequential dependencies are captured using position encodings that assume uniformly spaced, discrete tokens [ 41 ]. A similar strategy can be applied to GEOs by ordering them based on the time they were visited, providing an initial temporal context. However, visit order alone is insufficient to capture the rich temporal semantics of GEOs, particularly because the meaning of GEOs can change over time. For instance, a single location might function as a coffee shop in the morning and transition into a bar at night, reflecting distinct roles at different times of day. This suggests that GEO representations should be dynamic , adapting to temporal context inferred from mobility data. Designing such temporally adaptive representations remains an open challenge. While most prior work focuses on modeling trajectories as temporal sequences [ 14 , 20 , 22 ], relatively little attention has been paid to how the semantics of individual GEOs evolve over time. This raises a fundamental questions: Should a single GEO have multiple representations that vary across time? And if so, what should the temporal granularity of these representations be? Temporal behaviors in human mobility are often multi-scale, making it unclear how fine-grained these representations should be and how to aggregate them effectively. Answering these questions is critical for building models that treat time as an integral part of GEO representation, rather than as an auxiliary input. 2.3. The Vocabulary Challenge With GEOs identified and encoded, the subsequent challenge is to determine their representation for effective learning. In LLMs, each word or subword token is assigned a discrete ID from a fixed vocabulary, typically around 128k tokens in size [ 3 , 29 ]. While the specific tokenization technique matters, the decision principles are clear. One might think that we can similarly assign a unique identifier to each GEO, and learn a corresponding embedding. However, the space of possible objects on the map, is orders of magnitude larger, with hundreds of millions of locations worldwide, and follows an extremely long-tailed distribution. This is further complicated by the fact that some GEOs are frequently visited (e.g., airports, road segments), while others are rarely or never re-visited (e.g., private homes), leading to severe data sparsity for unique identifiers. Representing each GEO with a dedicated embedding is both computationally prohibitive and will yield poor generalization in data-sparse regions with few visitation patterns. More fundamentally, fixed vocabularies contrast with the continuous nature of geographic space, where new or rarely visited locations are constantly encountered. Some recent methods attempt to bypass discrete identifiers by embedding raw spatial and temporal signals directly [ 20 , 51 ], or by learning higher-level clusters to represent fine-grained spatial locations within graph neural networks, thereby improving scalability [ 43 ]. These initial attempts, though useful, are limited to specific GEOs and downstream tasks, thereby not fully capturing the complexity of geospatial semantics. This raises a fundamental question for mobility-based GEO modeling: How can we construct representations for a vast, sparse, and continuously evolving set of GEOs without relying on predefined vocabularies? 2.4. Hard(er) Constraints Unlike language, where any token can, in principle, appear in any position, modeling GEOs through mobility poses fundamental real-world constraints that must be respected to generate realistic representations. We discuss some of these these constraints below. 2.4.1. Accessibility and Reachability. Not all GEOs in a mobility sequence are equally accessible or reachable; a GPS trace cannot be arbitrarily associated with any GEO [ 21 ]. Physical access restrictions (e.g., private buildings, gated facilities), transportation constraints, and temporal feasibility (e.g., whether a location can be reached within a given time window) all affect which GEOs are plausible candidates. For example, a university campus may require an access pass, or a remote trailhead may be inaccessible without a vehicle. These constraints create a non-uniform feasibility landscape over geographic space, requiring models to reason not only about the locations a trajectory has visited, but also about which locations were realistically reachable given physical and temporal, and transportation mode constraints. 2.4.2. Capacity and Spatiotemporal Density Limits. Every GEO has intrinsic limits on how many agents can physically occupy or interact with it over space and time. A concert venue, for instance, cannot accommodate unlimited attendees regardless of demand. Similarly, a multi-story office tower can support far more occupants than a small neighborhood park, even if both have similar ground-level footprints. These capacity constraints are not merely operational considerations; they are fundamental semantic properties that influence how a GEO functions. Failing to account for these constraints can lead models to inaccurately assume that a GEO can support more activity than is physically or operationally feasible, resulting in unrealistic outputs in tasks such as demand forecasting, crowd simulation, or mobility prediction. Accurate GEO representations must therefore account for spatiotemporal density limits to support meaningful and physically plausible inference. 3. Potential Impact We envision GEO representations serving as a fundamental layer for geospatial foundation models (GEOFMs), enabling a wide range of applications across domains, which we group into three categories: Object-centric tasks involve reasoning about individual GEOs and their attributes. Examples include improving maps quality [ 9 ], like detecting missing or mislabeled POIs, inferring building functions from mobility patterns, identifying access points to large venues (e.g., stadium entrances), and correcting road connectivity errors (e.g., missing links, wrong one-way assignments). Another example is decision support, which includes recommending optimal locations for new businesses based on visitation patterns, estimating the capacity of facilities (e.g., determining parking space capacity), and assisting drivers with context-aware navigation, such as detecting likely entrances or drop-off points near a destination. Mobility-centric tasks involve understanding and optimizing movement patterns across space and time. Applications include dynamic traffic management based on real-time mobility data [ 38 ], optimizing delivery and service routes, forecasting logistics demand, analyzing commuter flows for transit planning, and identifying mobility bottlenecks or under-served areas in transportation networks. Population-level tasks involve aggregating GEO representations across users, time, and space to uncover macro-scale patterns. Applications include assessing mobility equity, identifying tourist activity patterns, estimating demand for public services such as healthcare or transit, monitoring urban growth and land use change [ 44 , 49 ], detecting disruptions during large events or disasters [ 4 , 50 ], and informing long-term decisions for infrastructure investment and public service allocation. By learning structured, multi-scale representations of space, time, and function, GeoFMs could unify these capabilities within a single, general-purpose framework. 4. Orthogonal Considerations 4.1. Privacy Privacy concerns are critical for GEO representation learning, given the reliance on human mobility data. Individual trajectories, even when anonymized, can often be re-identified through spatio-temporal modeling, posing risks of unintended disclosure. This is especially concerning when handling sensitive locations like private homes, hospitals, and places of worship. Unlike NLP tokens, which are abstract and generally unrelated to individuals, GEOs are grounded in real-world entities and often reflect personal routines. These risks raise an important design question: Should certain classes of GEOs, such as private residences, be represented at all? While including them may improve coverage, it not only introduces serious privacy vulnerabilities but also offers limited value for general-purpose tasks. Responsible representation learning may therefore require filtering, abstracting, or omitting sensitive GEOs altogether, alongside the use of privacy-preserving techniques such as differential privacy [ 2 , 17 ] or federated learning to enable decentralized model training without sharing raw trajectory data across devices [ 28 ]. Balancing representational utility with ethical safeguards is essential for deploying trustworthy GeoFMs in practice. 4.2. Cross Region Transferability Publicly available mobility data are typically restricted to specific geographic areas and narrow time spans, making it difficult to obtain comprehensive, large-scale coverage for training. As a result, GEO representations learned from such data risk being overly specialized to the regions and time periods they were derived from. Cross-region transferability is a core requirement for general-purpose GeoFMs. But transferring representations across regions is challenging due to substantial variation in mobility behavior, land use, and spatial semantics [ 10 , 23 , 50 ]. For example, the distribution of POIs, transportation modes, and urban density in Tokyo differs significantly from that in Los Angeles. Therefore, models trained on region-specific patterns may fail to generalize to areas with different structural or behavioral dynamics. To support such transferability, GEO representations must go beyond encoding region-specific mobility patterns. They should also encode spatial priors that capture differences in urban form (e.g., density, land use, connectivity), scale (e.g., city vs. neighborhood), and mobility modality (e.g., walking vs. driving), and that can adapt to distributional shifts both in the physical layout of geographic space and in the mobility behaviors associated with it. 4.3. Interpretability As GEO representations grow in dimensionality, they get harder to understand. Yet interpretability remains essential, especially in high-stakes domains such as urban planning, transportation, and public policy, where stakeholders must be able to understand and justify spatial decisions or model-driven recommendations. In language modeling, word embeddings have been shown to align with interpretable semantic axes, such as gender, tense, or country-capital relationships. Similarly, GEO embeddings should aim to reveal dimensions that correspond to meaningful spatio-temporal and contextual attributes [ 12 ], but this level of interpretability remains largely underexplored in current research. Improving interpretability in GEO representations may benefit from techniques adapted from language models. For example, embedding probes can be repurposed to test whether GEO embeddings capture meaningful attributes such as accessibility or population density. Overall, this is a promising direction for building more transparent and accountable GeoFMs, particularly in applications where explainable decision-making is critical. 5. Conclusion In this paper, we introduced GEOs as a unifying abstraction for representing geospatial objects and outlined the core challenges in learning their representations across spatial, contextual, and temporal dimensions. We highlighted the unique technical difficulties of modeling GEOs using human mobility data, including issues like sparsity, scale, transferability, privacy and interpretability. We also explained why GEOs require fundamentally different modeling assumptions than language tokens. We argue that GEOs represent a critical building block for general-purpose GeoFMs, and achieving this goal requires collaborative efforts from the community. Figure 1: Open in a new tab Pipeline overview for mobility-enriched GEOs: approach, applications, and modeling considerations. CCS Concepts. • Computing methodologies → Neural networks; • Information systems → Geographic information systems. Acknowledgments This research has been funded in part by the NIH award R01LM014026, and NSF awards CNS-2125530 and DMS-2428039. References [1]. Agarwal Mohit, Sun Mimi, Kamath Chaitanya, Muslim Arbaaz, Sarker Prithul, Paul Joydeep, Yee Hector, Sieniek Marcin, Jablonski Kim, Mayer Yael, et al. 2024. General Geospatial Inference with a Population Dynamics Foundation Model. arXiv preprint arXiv:2411.07207 (2024). [ Google Scholar ] [2]. Ahuja Ritesh, Zeighami Sepanta, Ghinita Gabriel, and Shahabi Cyrus. 2023. A neural approach to spatio-temporal data release with user-level differential privacy. Proceedings of the ACM on Management of Data 1, 1 (2023), 1–25. [ Google Scholar ] [3]. Ali Mehdi, Fromm Michael, Thellmann Klaudia, Rutmann Richard, Lübbering Max, Leveling Johannes, Klug Katrin, Ebert Jan, Doll Niclas, Buschhoff Jasper, et al. 2024. Tokenizer choice for llm training: Negligible or crucial?. In Findings of the Association for Computational Linguistics: NAACL 2024. 3907–3924. [ Google Scholar ] [4]. Azarijoo Bita, Siampou Maria Despoina, Krumm John, and Shahabi Cyrus. 2025. ICAD: A Self-Supervised Autoregressive Approach for Multi-Context Anomaly Detection in Human Mobility Data. In Proceedings of the 33rd ACM International Conference on Advances in Geographic Information Systems. Accepted for publication. [ Google Scholar ] [5]. Balsebre Pasquale, Huang Weiming, Cong Gao, and Li Yi. 2024. City foundation models for learning general purpose representations from openstreetmap. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 87–97. [ Google Scholar ] [6]. Bommasani Rishi, Hudson Drew A, Adeli Ehsan, Altman Russ, Arora Simran, von Arx Sydney, Bernstein Michael S, Bohg Jeannette, Bosselut Antoine, Brunskill Emma, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021). [ Google Scholar ] [7]. Cheng Jiawei, Wang Jingyuan, Zhang Yichuan, Ji Jiahao, Zhu Yuanshao, Zhang Zhibo, and Zhao Xiangyu. 2025. POI-Enhancer: An LLM-based Semantic Enhancement Framework for POI Representation Learning. In Proceedings of the AAAI conference on artificial intelligence, Vol. 39. 11509–11517. [ Google Scholar ] [8]. Choudhury Shushman, Aharoni Elad, Suvarna Chandrakumari, Tsogsuren Iveel, Kreidieh Abdul Rahman, Lu Chun-Ta, and Arora Neha. 2025. S2Vec: Self-Supervised Geospatial Embeddings. arXiv preprint arXiv:2504.16942 (2025). [ Google Scholar ] [9]. Choudhury Shushman, Kreidieh Abdul Rahman, Kuznetsov Ivan, and Arora Neha. 2024. Towards a Trajectory-powered Foundation Model of Mobility. In Proceedings of the 3rd ACM SIGSPATIAL International Workshop on Spatial Big Data and AI for Industrial Applications. 1–4. [ Google Scholar ] [10]. Chu Chen, Shahabi Cyrus, Tung Emmanuel, and Shafique Khurram. 2025. One Model, Many Cities: A Transferable Social Relationship Inference Framework for Human Mobility Data. In Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems (SIGSPATIAL ‘25). Association for Computing Machinery, New York, NY, USA, 11 pages. doi: 10.1145/3748636.3762710 [ DOI ] [ Google Scholar ] [11]. Gao Yunfan, Xiong Yun, Wang Siqi, and Wang Haofen. 2022. Geobert: Pretraining geospatial representation learning on point-of-interest. Applied Sciences 12, 24 (2022), 12942. [ Google Scholar ] [12]. Gurnee Wes and Tegmark Max. 2024. Language Models Represent Space and Time. In The Twelfth International Conference on Learning Representations. [ Google Scholar ] [13]. Hajisafi Arash, Lin Haowen, Shaham Sina, Hu Haoji, Siampou Maria Despoina, Chiang Yao-Yi, and Shahabi Cyrus. 2023. Learning dynamic graphs from all contextual information for accurate point-of-interest visit forecasting. In Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems. 1–12. [ Google Scholar ] [14]. Hsu Shang-Ling, Tung Emmanuel, Krumm John, Shahabi Cyrus, and Shafique Khurram. 2024. Trajgpt: Controlled synthetic trajectory generation using a multitask transformer-based spatiotemporal model. In Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems. 362–371. [ Google Scholar ] [15]. Janowicz Krzysztof, Gao Song, McKenzie Grant, Hu Yingjie, and Bhaduri Budhendra. 2020. GeoAI: spatially explicit artificial intelligence techniques for geographic knowledge discovery and beyond. 625–636 pages. [ Google Scholar ] [16]. Klemmer Konstantin, Safir Nathan S, and Neill Daniel B. 2023. Positional encoder graph neural networks for geographic data. In International conference on artificial intelligence and statistics. PMLR, 1379–1389. [ Google Scholar ] [17]. Krumm John. 2007. Inference attacks on location tracks. In International Conference on Pervasive Computing. Springer, 127–143. [ Google Scholar ] [18]. Li Yi, Huang Weiming, Cong Gao, Wang Hao, and Wang Zheng. 2023. Urban region representation learning with openstreetmap building footprints. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1363–1373. [ Google Scholar ] [19]. Li Zekun, Kim Jina, Chiang Yao-Yi, and Chen Muhao. 2022. Spabert: a pretrained language model from geographic data for geo-entity representation. arXiv preprint arXiv:2210.12213 (2022). [ Google Scholar ] [20]. Lin Haowen, Chiang Yao-Yi, Xiong Li, and Shahabi Cyrus. 2024. Unified Modeling and Clustering of Mobility Trajectories with Spatiotemporal Point Processes. In Proceedings of the 2024 SIAM International Conference on Data Mining (SDM). SIAM, 625–633. [ Google Scholar ] [21]. Lin Haowen, Krumm John, Shahabi Cyrus, and Xiong Li. 2024. Controllable Visit Trajectory Generation with Spatiotemporal Constraints. In 2024 IEEE International Conference on Data Mining (ICDM). IEEE, 773–778. [ Google Scholar ] [22]. Lin Yan, Wan Huaiyu, Guo Shengnan, and Lin Youfang. 2021. Pre-training context and time aware location embeddings from spatial-temporal trajectories for user next location prediction. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 4241–4248. [ Google Scholar ] [23]. Lin Yan, Wei Tonglong, Zhou Zeyu, Wen Haomin, Hu Jilin, Guo Shengnan, Lin Youfang, and Wan Huaiyu. 2024. TrajFM: A vehicle trajectory foundation model for region and task transferability. arXiv preprint arXiv:2408.15251 (2024). [ Google Scholar ] [24]. Mai Gengchen, Huang Weiming, Sun Jin, Song Suhang, Mishra Deepak, Liu Ninghao, Gao Song, Liu Tianming, Cong Gao, Hu Yingjie, et al. 2023. On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798 (2023). [ Google Scholar ] [25]. Mai Gengchen, Janowicz Krzysztof, Hu Yingjie, Gao Song, Yan Bo, Zhu Rui, Cai Ling, and Lao Ni. 2022. A review of location encoding for GeoAI: methods and applications. International Journal of Geographical Information Science 36, 4 (2022), 639–673. [ Google Scholar ] [26]. Mai Gengchen, Janowicz Krzysztof, Yan Bo, Zhu Rui, Cai Ling, and Lao Ni. 2020. Multi-scale representation learning for spatial feature distributions using grid cells. arXiv preprint arXiv:2003.00824 (2020). [ Google Scholar ] [27]. Mai Gengchen, Jiang Chiyu, Sun Weiwei, Zhu Rui, Xuan Yao, Cai Ling, Janowicz Krzysztof, Ermon Stefano, and Lao Ni. 2023. Towards general-purpose representation learning of polygonal geometries. GeoInformatica 27, 2 (2023), 289–340. [ Google Scholar ] [28]. Meng Chuizheng, Rambhatla Sirisha, and Liu Yan. 2021. Cross-node federated graph neural network for spatio-temporal data modeling. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 1202–1211. [ Google Scholar ] [29]. Mielke Sabrina J, Alyafeai Zaid, Salesky Elizabeth, Raffel Colin, Dey Manan, Gallé Matthias, Raja Arun, Si Chenglei, Lee Wilson Y, Sagot Benoît, et al. 2021. Between words and characters: A brief history of open-vocabulary modeling and tokenization in NLP. arXiv preprint arXiv:2112.10508 (2021). [ Google Scholar ] [30]. Mikolov Tomas, Chen Kai, Corrado Greg, and Dean Jeffrey. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013). [ Google Scholar ] [31]. Momennejad Ida, Hasanbeig Hosein, Frujeri Felipe Vieira, Sharma Hiteshi, Jojic Nebojsa, Palangi Hamid, Ness Robert, and Larson Jonathan. 2023. Evaluating cognitive maps and planning in large language models with cogeval. Advances in Neural Information Processing Systems 36 (2023), 69736–69751. [ Google Scholar ] [32]. Musleh Mashaal and Mokbel Mohamed F. 2024. Let’s Speak Trajectories: A Vision to Use NLP Models for Trajectory Analysis Tasks. ACM Transactions on Spatial Algorithms and Systems 10, 2 (2024), 1–25. [ Google Scholar ] [33]. Musleh Mashaal, Mokbel Mohamed F, and Abbar Sofiane. 2022. Let’s speak trajectories. In Proceedings of the 30th International Conference on Advances in Geographic Information Systems. 1–4. [ Google Scholar ] [34]. Newson Paul and Krumm John. 2009. Hidden Markov map matching through noise and sparseness. In Proceedings of the 17th ACM SIGSPATIAL international conference on advances in geographic information systems. 336–343. [ Google Scholar ] [35]. Nishida Kyosuke, Toda Hiroyuki, Kurashima Takeshi, and Suhara Yoshihiko. 2014. Probabilistic identification of visited point-of-interest for personalized automatic check-in. In Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing. 631–642. [ Google Scholar ] [36]. SafeGraph. 2022. Determining Points of Interest Visits From Location Data: A Technical Guide To Visit Attribution. https://www.safegraph.com/guides/visitattribution-white-paper . Accessed: 2025-05-13. [37]. Saxena Nripsuta Ani, Hsu Shang-Ling, Shetty Mehul, Alkhadra Omar, Shahabi Cyrus, and Horn Abigail L. 2025. POIFormer: A Transformer-Based Framework for Accurate and Scalable Point-of-Interest Attribution. arXiv preprint arXiv:2507.09137 (2025). [ Google Scholar ] [38]. Siampou Maria Despoina, Anastasiou Chrysovalantis, Krumm John, and Shahabi Cyrus. 2025. TrajRoute: Rethinking Routing with a Simple Trajectory-Based Approach–Forget the Maps and Traffic!. In 2025 26th IEEE International Conference on Mobile Data Management (MDM). IEEE, 168–174. [ Google Scholar ] [39]. Siampou Maria Despoina, Li Jialiang, Krumm John, Shahabi Cyrus, and Lu Hua. 2025. Poly2Vec: Polymorphic Fourier-Based Encoding of Geospatial Objects for GeoAI Applications. In Forty-second International Conference on Machine Learning. [ Google Scholar ] [40]. Suzuki Jun, Suhara Yoshihiko, Toda Hiroyuki, and Nishida Kyosuke. 2019. Personalized visited-poi assignment to individual raw GPS trajectories. ACM Transactions on Spatial Algorithms and Systems (TSAS) 5, 3 (2019), 1–28. [ Google Scholar ] [41]. Vaswani Ashish, Shazeer Noam, Parmar Niki, Uszkoreit Jakob, Jones Llion, Gomez Aidan N, Kaiser Łukasz, and Polosukhin Illia. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017). [ Google Scholar ] [42]. Wang Zhecheng, Li Haoyuan, and Rajagopal Ram. 2020. Urban2vec: Incorporating street view imagery and pois for multi-modal urban neighborhood embedding. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 1013–1020. [ Google Scholar ] [43]. Wölker Yannick, Hajisafi Arash, Shahabi Cyrus, and Renz Matthias. 2025. Small Graph Is All You Need: DeepStateGNN for Scalable Traffic Forecasting. arXiv preprint arXiv:2502.14525 (2025). [ Google Scholar ] [44]. Wu Shangbin, Yan Xu, Fan Xiaoliang, Pan Shirui, Zhu Shichao, Zheng Chuanpan, Cheng Ming, and Wang Cheng. 2022. Multi-graph fusion networks for urban region embedding. arXiv preprint arXiv:2201.09760 (2022). [ Google Scholar ] [45]. Xiao Congxi, Zhou Jingbo, Xiao Yixiong, Huang Jizhou, and Xiong Hui. 2024. ReFound: Crafting a Foundation Model for Urban Region Understanding upon Language and Visual Foundations. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3527–3538. [ Google Scholar ] [46]. Yamada Yutaro, Bao Yihan, Lampinen Andrew Kyle, Kasai Jungo, and Yildirim Ilker. 2024. Evaluating Spatial Understanding of Large Language Models. Transactions on Machine Learning Research (2024). [ Google Scholar ] [47]. Yu Dazhou, Hu Yuntong, Li Yun, and Zhao Liang. 2024. PolygonGNN: Representation learning for polygonal geometries with heterogeneous visibility graph. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4012–4022. [ Google Scholar ] [48]. Yuan Yuan, Ding Jingtao, Feng Jie, Jin Depeng, and Li Yong. 2024. Unist: A prompt-empowered universal model for urban spatio-temporal prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4095–4106. [ Google Scholar ] [49]. Zhang Mingyang, Li Tong, Li Yong, and Hui Pan. 2021. Multi-view joint graph representation learning for urban region embedding. In Proceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence. 4431–4437. [ Google Scholar ] [50]. Zhang Zheng, Amiri Hossein, Yu Dazhou, Hu Yuntong, Zhao Liang, and Andreas Züfle. 2024. Transferable Unsupervised Outlier Detection Framework for Human Semantic Trajectories. In Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems. 350–360. [ Google Scholar ] [51]. Zhu Yuanshao, Yu James Jianqiao, Zhao Xiangyu, Wei Xuetao, and Liang Yuxuan. 2024. UniTraj: Universal human trajectory modeling from billion-scale worldwide traces. arXiv preprint arXiv:2411.03859 (2024). [ Google Scholar ] ACTIONS View on publisher site PDF (443.5 KB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top