Attentive Graph Enhanced Region Representation Learning

TL;DR

This work tackles urban region representation by learning robust embeddings from heterogeneous data sources. It introduces ATGRL, a three-component framework consisting of graph-enhanced learning (mobility, function, semantics graphs with noise filtering), multi-graph aggregation (cosine-based attention with multi-heads), and a dual-stage fusion (linear attention and gated fusion) to capture local and global dependencies. The model is trained with origin-destination, function, and semantics predictors, achieving superior performance on check-in prediction, land-use classification, and crime prediction while improving computational efficiency. The results indicate significant gains over state-of-the-art baselines and demonstrate ATGRL’s practical value for urban planning and analysis, with scalable design suitable for real-world data.

Abstract

Representing urban regions accurately and comprehensively is essential for various urban planning and analysis tasks. Recently, with the expansion of the city, modeling long-range spatial dependencies with multiple data sources plays an important role in urban region representation. In this paper, we propose the Attentive Graph Enhanced Region Representation Learning (ATGRL) model, which aims to capture comprehensive dependencies from multiple graphs and learn rich semantic representations of urban regions. Specifically, we propose a graph-enhanced learning module to construct regional graphs by incorporating mobility flow patterns, point of interests (POIs) functions, and check-in semantics with noise filtering. Then, we present a multi-graph aggregation module to capture both local and global spatial dependencies between regions by integrating information from multiple graphs. In addition, we design a dual-stage fusion module to facilitate information sharing between different views and efficiently fuse multi-view representations for urban region embedding using an improved linear attention mechanism. Finally, extensive experiments on real-world datasets for three downstream tasks demonstrate the superior performance of our model compared to state-of-the-art methods.

Figures (7)

• Figure 1: The Borough of Manhattan in New York City serves as the study area for this paper. As shown in sub-figure 1, the Community Boards Berg__New have divided Manhattan into 12 districts based on land use. To conduct a more comprehensive analysis, the study area is divided into 180 regions.
• Figure 2: The architecture of ATGRL consists of three components: (A) The graph-enhanced learning module for constructing multi-graph from features extracted from human mobility data, functional features, and semantic features and filtering the graphs for noise using soft thresholding. (B) The multi-graph Aggregation Module for capturing non-linear and long-range dependencies between regions using improved cosine similarity graph attention mechanism to generate corresponding region representations. (C) The dual-stage fusion module utilizes an improved linear attention mechanism to achieve information sharing between different views and efficiently fuses multi-view representations for urban region embedding.
• Figure 3: Illustration of (a) soft thresholding and (b) its derivative.
• Figure 4: Ablation studies for two tasks on NYC dataset.
• Figure 5: Districts in Manhattan and region clusters.

Theorems & Definitions (4)

• Definition 1: Human mobility feature
• Definition 2: Function feature
• Definition 3: Semantics feature
• Definition 4: Urban region embedding problem