Abstract
In recent years, origin-destination (OD) demand prediction has gained significant attention for its profound implications in urban development. Existing data-driven deep learning methods primarily focus on the spatial or temporal dependency between regions yet neglecting regions' fundamental functional difference. Though knowledge-driven physical methods have characterised regions' functions by their radiation and attraction capacities, these functions are defined on numerical factors like population without considering regions' intrinsic nominal attributes, e.g., a region is a residential or industrial district. Moreover, the complicated relationships between two types of capacities, e.g., the radiation capacity of a residential district in the morning will be transformed into the attraction capacity in the evening, are totally missing from physical methods. In this paper, we not only generalize the physical radiation and attraction capacities into the deep learning framework with the extended capability to fulfil regions' functions, but also present a new model that captures the relationships between two types of capacities. Specifically, we first model regions' radiation and attraction capacities using a bilateral branch network, each equipped with regions' attribute representations. We then describe the transformation relationship of different capacities of the same region using a hypergraph-based parameter generation method. We finally unveil the competition relationship of different regions with the same attraction capacity through cluster-based adversarial learning. Extensive experiments on two datasets demonstrate the consistent improvements of our method over the state-of-the-art baselines, as well as the good explainability of regions' functions using their nominal attributes.
Introduction
Predicting Origin-Destination (OD) demand—the volume of trips between specific urban locations—is a critical task for ride-hailing platforms and urban planners. While modern deep learning models excel at identifying patterns in large-scale data, they often treat city blocks as generic nodes in a graph, overlooking the fundamental functional differences between regions. For instance, a residential district behaves fundamentally differently from an industrial park or a commercial hub.
Traditional physical models, such as the Gravity Model, characterize these differences through "radiation" (outflow) and "attraction" (inflow) capacities. However, these models typically rely on simple numerical factors like population and distance, failing to capture "nominal" attributes (e.g., land use types like "theater district") or the complex temporal transformations of these roles. A residential area may act as a radiation source in the morning but transform into an attraction destination in the evening.
The RACTC Framework
To bridge the gap between physical modeling and deep learning, we propose the RACTC (Radiation and Attraction with Transformation and Competition) framework. This architecture integrates the physical concepts of radiation and attraction into a high-performance neural network.
Bilateral Branch Network
The model employs two separate branches to learn distinct representations for origins (radiation) and destinations (attraction). This structure acknowledges that the features driving a person to leave a region are not identical to those drawing them to one.
Attribute Hypergraphs and Transformation
Standard graphs are limited to pairwise connections. RACTC utilizes attribute hypergraphs to model "many-to-many" relationships between regions and their Points of Interest (POIs). This captures the unique functional "flavor" of a neighborhood. Furthermore, a hypergraph-based parameter generation method learns the transformation relationship, modeling how a region's capacity shifts over time—for example, how a theater district's attraction peaks before a show and transforms into radiation afterward.
Competition and Population Enhancement
The framework explicitly models competition relationships between similar regions (e.g., two nearby restaurant hubs) using cluster-based adversarial learning with a Gradient Reversal Layer. This helps the model distinguish between competing destinations rather than just identifying similarity. Additionally, population data is integrated via a logistic distribution to reinforce the radiation branch, grounding the model in the physical reality that higher populations typically generate more trips.
Experimental Evaluation
The model was evaluated on two real-world datasets: the New York Taxi Dataset (Manhattan) and the Chicago Taxi Dataset. Trip data was enriched with geospatial data from OpenStreetMap and demographic data from the US Census Bureau.
Key Results
RACTC consistently outperformed state-of-the-art baselines, including GEML, MPGCN, and CMOD:
- Accuracy: RACTC achieved a 10.17% improvement in MAE on the New York dataset and a 13.02% improvement on the Chicago dataset compared to the best baseline.
- Explainability: Attention mechanisms revealed that specific POI attributes (e.g., "place of worship," "sports center") drive demand at predictable times, aligning with real-world urban behavior.
- Ablation Studies: Removing the bilateral branches or hypergraph components led to significant performance drops, validating the radiation/attraction perspective.
Conclusion
RACTC demonstrates that integrating physical urban modeling concepts into deep learning frameworks significantly enhances OD demand prediction. By treating regions as dynamic actors with time-sensitive functional roles, the model achieves superior accuracy and explainability. Future work may explore extending this radiation/attraction perspective to other urban flows, such as bike-sharing or digital information spread, and addressing the sensitivity of clustering parameters in competition modeling.
Summary updated