Introduction
Predicting citywide crowd flow—the movement of people across urban grids—is essential for modern smart city management. While Convolutional Neural Networks (CNNs) have become the standard for this "raster-based" prediction task, they suffer from two critical limitations that hinder their performance in complex urban environments.
First, standard CNNs struggle with inefficient global dependency modeling. To capture how a distant residential area influences a central business district, models typically require stacking numerous layers to expand the receptive field, leading to high computational costs and training difficulties.
Second, existing models often employ a shared predictor, applying identical weights across all grid cells. This "one-size-fits-all" approach ignores the fact that different urban regions—such as parks, schools, and offices—exhibit fundamentally different flow patterns.
The DeepLGR Framework
To address these challenges, we introduce DeepLGR, a novel framework composed of three specialized modules:
- Local Feature Extraction (L): Utilizing Squeeze-and-Excitation Networks (SENet), the model adaptively weights input channels. This is particularly effective for processing temporal data where recent, daily, and weekly trends are represented as distinct channels.
- Global Context Module (G): Instead of deep stacking, DeepLGR uses Spatial Pyramid Pooling (SPP) to capture multi-scale global priors (e.g., $1\times1, 2\times2, 4\times4$). These are upsampled using Subpixel blocks to maintain spatial integrity while providing a citywide perspective efficiently.
- Region-Specific Predictor (R): To allow for localized prediction logic without a massive parameter increase, we employ Tucker Decomposition. By factorizing the weight tensor, the model learns unique behaviors for every grid cell, preserving the Euclidean structure of the city while remaining lightweight.
Experimental Evaluation
DeepLGR was benchmarked against state-of-the-art models using two real-world datasets:
- TaxiBJ: A large-scale taxi flow dataset in Beijing ($128 \times 128$ grid).
- HappyValley: Human flow data from a Beijing theme park, characterized by highly skewed density distributions.
Analysis and Results
The core insight of DeepLGR is that urban regions are heterogeneous. As shown in the data distributions below, different functional zones (office, residence, park) have distinct temporal signatures and flow densities.
By accounting for these regional differences and global contexts, DeepLGR achieves significant performance gains:
- Superior Accuracy: On TaxiBJ, it achieved a 10.2% improvement in MAE and a 44.1% improvement in SMAPE over previous leaders like DeepSTN+.
- High Efficiency: On the HappyValley dataset, DeepLGR outperformed competitors while using only 6.2% of the parameters (0.97M vs. 15.70M).
- Faster Convergence: The model requires roughly half the training iterations of earlier frameworks like UrbanFM.
Conclusion and Future Directions
DeepLGR demonstrates that efficient global modeling and region-specific customization are key to advancing urban flow analytics. While currently optimized for grid-based data, future work will explore applying the "Local-Global-Region" philosophy to Graph Convolutional Networks (GCNs) to better model non-grid structures like road and subway networks. Additionally, Neural Architecture Search (NAS) could further automate the optimization of these components for diverse city layouts.
Summary updated