TransFlower: An Explainable Transformer-Based Model with Flow-to-Flow Attention for Commuting Flow Prediction

TL;DR

TransFlower tackles commuting flow prediction by combining a geo-spatial encoder with a transformer-based flow predictor and a flow-to-flow attention mechanism. It introduces an anisotropy-aware relative location encoder to capture spatial directionality alongside distance, and demonstrates that modeling interactions among flows yields superior predictive performance. Across CA, MA, and TX datasets, it achieves up to 30.8% CPC improvement over strong baselines and provides interpretable explanations through attention maps and location-embedding clustering. The work offers a practical, explainable framework for urban planning that links geographic features to mobility dynamics with potential relevance to SDGs.

Abstract

Understanding the link between urban planning and commuting flows is crucial for guiding urban development and policymaking. This research, bridging computer science and urban studies, addresses the challenge of integrating these fields with their distinct focuses. Traditional urban studies methods, like the gravity and radiation models, often underperform in complex scenarios due to their limited handling of multiple variables and reliance on overly simplistic and unrealistic assumptions, such as spatial isotropy. While deep learning models offer improved accuracy, their black-box nature poses a trade-off between performance and explainability -- both vital for analyzing complex societal phenomena like commuting flows. To address this, we introduce TransFlower, an explainable, transformer-based model employing flow-to-flow attention to predict urban commuting patterns. It features a geospatial encoder with an anisotropy-aware relative location encoder for nuanced flow representation. Following this, the transformer-based flow predictor enhances this by leveraging attention mechanisms to efficiently capture flow interactions. Our model outperforms existing methods by up to 30.8% Common Part of Commuters, offering insights into mobility dynamics crucial for urban planning and policy decisions.

Figures (4)

• Figure 1: The framework of TransFlower which contains two major components: geo-spatial encoder and flower predictor.
• Figure 2: Comparison between different modules.
• Figure 3: Location embedding clustering of the relative location encoder from the trained TransFlower.
• Figure 4: Case study for flow prediction of one origin region in San Francisco.