RegionLM: Integrating Point, Line, and Polygon Context for Urban Representation Learning
July 15, 2026 11:00 am (Central Time)
Abstract
RegionLM is a geospatial representation learning pipeline for generating contextual region embeddings from heterogeneous OpenStreetMap (OSM) features, including points, lines, and polygons. Built on SpaBERT-style spatial language modeling, RegionLM extracts features within target regions, converts nearby spatial context into pseudo-sentences containing feature semantics and geographic relationships, and learns contextual embeddings that jointly encode semantic, spatial, and topological information. The framework rasterizes line and polygon features into shared spatial grids, aggregates feature-level embeddings into region-level representations, and clusters regions into contextual categories that capture neighborhood structure and urban function. RegionLM supports end-to-end urban representation learning for downstream applications such as mobility modeling and trajectory anomaly detection.
Speakers
Yao-Yi Chiang
University of Minnesota
Yao-Yi Chiang is a Professor in the Computer Science & Engineering Department at the University of Minnesota. Previously, he was an Associate Professor (Research) in Spatial Sciences at the University of Southern California. Dr. Chiang's research interests are in spatial artificial intelligence. He develops machine learning methods to understand complex environmental phenomena and human-environment interactions using multimodal data that can be sparse, unevenly distributed, and span varying spatiotemporal scales. Dr. Chiang has received funding from various government agencies, including NSF, NEH, NIH, DARPA, IARPA, and NGA, as well as industry partners such as NTT Global Networks and BAE Systems. He was a visiting researcher at Google AI, a machine learning consultant at Meta, and the chief scientist at AirMap. Dr. Chiang founded Kartta Foundation, a non-profit organization that provides software and services to refine and assemble geographic knowledge for the public good. Kartta Foundation manages Kartta Labs, a previous Google product.
Yijun Lin
Southern Illinois University Carbondale
Yijun Lin is an incoming Assistant Professor in the School of Computing at Southern Illinois University Carbondale. She is currently working with Dr. Yao-Yi Chiang at the University of Minnesota and is expected to complete her degree in June 2026. Her research lies in the areas of AI, machine learning, and spatial science. She is a recipient of the University of Minnesota Doctoral Dissertation Fellowship and the Data Science Initiative ADC Fellowship. Her work has been published in leading venues including ACM SIGKDD, ACM SIGSPATIAL, and IEEE ICDM. She is also active in professional service, including serving as a lead organizer of the ICDAR MapText Competition, PC Chair of the ACM SIGSPATIAL Workshop on Human-Centered Geospatial Computing, and PC for conferences including SIGKDD, SIGSPATIAL, and SDM.
Zekun Li
University of Minnesota
Zekun Li is a computer science Ph.D. student at the University of Minnesota (UMN). Her research interests lie in the automatic understanding of historical maps using computer vision and natural language processing techniques. She has engaged in work involving the detection of text on historical map labels, the connection of separated text labels, the linking of recognized place names to existing knowledge bases (entity linking), and the inference of label types (entity typing).