SARN: Structurally-Aware Recurrent Network for Spatio-Temporal Disaggregation

TL;DR

The paper addresses disaggregating spatio-temporal data from coarse, irregular partitions to fine-grained regions by introducing SARN, a GRU-based model augmented with Structurally-Aware Spatial Attention (SASA) that uses global and containment-guided structural attention to capture cross-level spatial interactions. It demonstrates that SASA-enabled learning yields consistent improvements over neural baselines and heuristics across NYC taxi and bikeshare datasets, with faster convergence and robust transfer learning capabilities under data-scarce conditions. Additional contributions include showing that disaggregated data are authentic for downstream traffic prediction and can support realistic synthesis of individual events, accompanied by a practical transfer-learning protocol that reduces training time from hours to mere minutes. The work offers a scalable approach for privacy-preserving, high-resolution urban analytics and suggests a broad, reusable framework for cross-level spatio-temporal modeling in urban AI.

Abstract

Open data is frequently released spatially aggregated, usually to comply with privacy policies. But coarse, heterogeneous aggregations complicate learning and integration for downstream AI/ML systems. In this work, we consider models to disaggregate spatio-temporal data from a low-resolution, irregular partition (e.g., census tract) to a high-resolution, irregular partition (e.g., city block). We propose an overarching model named the Structurally-Aware Recurrent Network (SARN), which integrates structurally-aware spatial attention (SASA) layers into the Gated Recurrent Unit (GRU) model. The spatial attention layers capture spatial interactions among regions, while the gated recurrent module captures the temporal dependencies. Each SASA layer calculates both global and structural attention -- global attention facilitates comprehensive interactions between different geographic levels, while structural attention leverages the containment relationship between different geographic levels (e.g., a city block being wholly contained within a census tract) to ensure coherent and consistent results. For scenarios with limited historical training data, we explore transfer learning and show that a model pre-trained on one city variable can be fine-tuned for another city variable using only a few hundred samples. Evaluating these techniques on two mobility datasets, we find that on both datasets, SARN significantly outperforms other neural models (5% and 1%) and typical heuristic methods (40% and 14%), enabling us to generate realistic, high-quality fine-grained data for downstream applications.

Figures (8)

• Figure 1: Disaggregation problem setting and its utilities. Our proposed Structurally-Aware Recurrent Network (SARN) is trained with coarse, aggregated data and structural containment maps. The trained model can be utilized in downstream applications, such as traffic prediction. It also can be adopted to disaggregate another variable through weight-sharing and fine-tuning with only a limited amount of data.
• Figure 2: Four distinct geographic aggregation levels for the same NYC area. PUMA regions are the largest and most coarse. BLOCK regions are smallest and finest.
• Figure 3: Model architecture of our proposed SARN. (a) Original architecture of the GRU model. Fully-connected layers (FC) are applied on input $X^i$. (b) Our proposed SARN architecture. The fully-connected layers are replaced with our proposed structurally-aware spatial attention (SASA) layers. (c) Our proposed SASA layer. A SASA layer is composed with multiple spatial attention heads. Each spatial attention head calculates global and structural attention across regions from two different geographic levels. The containment map is incorporated to reflect the real-world structural hierarchy between geographic levels, and ensures the generation of coherent values across different resolutions.
• Figure 4: (a) Containment relationships between geographic level $g^i$ and $g^j$. Each region from $g^i$ is wholly contained in a region from $g^j$. (b) Binary containment map $M^{i,j}$ between georaphic level $g^i$ and $g^j$. (c) Structural-aware spatial attention is calculated among connected regions from two levels.
• Figure 5: Convergence plots of neural models. Top three plots: show NYC Taxi dataset. Bottom three plots: shows NYC Bikeshare dataset. Red line represents our proposed SARN model, and green line indicates the best results. In five out of six tasks, our SARN converged faster than other neural methods.