Spatiality Preservable Factored Poisson Regression for Large-Scale Fine-Grained GPS-Based Population Analysis

Masamichi Shimosaka, Yuta Hayakawa, Kota TsubouchiAAAI 2019 · DOI

Introduction

Analyzing urban population dynamics—the movement and density of people across a city—is fundamental for urban planning, disaster response, and commercial site selection. While traditional models perform adequately at coarse resolutions (e.g., 1km meshes), they often fail at the "fine-grained" level required to understand crowds at specific train stations or commercial buildings.

The challenge of scale: visualizing the 100m x 100m mesh resolution against a 3000m area.

At a 100m resolution, data becomes extremely sparse; many grid cells contain few to no GPS pings. This sparsity causes standard discriminative models to overfit and increases computational complexity exponentially as the number of meshes grows. This research introduces Spatiality Preservable Factored Regression (SPFR), a model designed to overcome these hurdles by leveraging shared behavioral patterns and spatial constraints.

Methodology

The SPFR framework is built upon three core technical pillars:

  1. Bilinear Poisson Regression: Given that population counts are non-negative integers, the model utilizes a Poisson distribution. The bilinear structure allows it to capture complex interactions between temporal features (e.g., time of day) and external factors (e.g., weather or day of the week).
  2. Parameter Factorization: To combat the "curse of dimensionality," SPFR assumes that the city is composed of a few latent "base patterns" (e.g., residential, business, or leisure behaviors). Instead of training unique parameters for every 100m square, each mesh is modeled as a mixture of these shared patterns, significantly reducing the parameter space.
  3. Graph Proximity Regularization: To maintain geographic consistency, the model incorporates a graph-based regularizer. This ensures that neighboring meshes—or meshes connected by infrastructure like railways—exhibit similar mixtures of patterns, preventing erratic predictions in sparse areas.

Experimental Setup

The model was evaluated using a massive dataset of anonymized GPS logs from a commercial mobile app in Japan. The data spanned 365 days (July 2013 – June 2014) in the Kanto (Tokyo) region, comprising approximately 15 million records per day. The researchers tested SPFR across multiple resolutions, focusing on the challenging 100m mesh scale.

Key Insights and Results

SPFR addresses the "cold start" problem in urban analysis—predicting behavior in low-traffic areas by "borrowing" information from well-traveled areas with similar latent patterns.

Visualization of a learned latent pattern mixture across the Tokyo metropolitan area.

Performance and Scalability

  • Accuracy: SPFR consistently outperformed standard bilinear Poisson models in Mean Negative Log Likelihood (MNLL), with the performance gap widening as the mesh size decreased to 100m.
  • Robustness: While baseline models degraded sharply due to data thinning at high resolutions, SPFR remained stable, demonstrating the effectiveness of its factorization and regularization components.
  • Interpretability: By applying simplex constraints to the mixture weights, the model allows for clear visualization of urban behaviors. Researchers could identify distinct patterns, such as the "two-peak" weekday commute versus the "mid-day peak" weekend leisure behavior, and map them to specific geographic districts.

Conclusion

By combining the strengths of factorization with spatial graph regularization, SPFR provides a scalable and accurate solution for fine-grained population analysis. Future work aims to incorporate dynamic graphs to account for real-time traffic events and to extend the model's application to anomaly detection, such as identifying unexpected crowds during emergencies.

Summary updated

BibTeX
@inproceedings{shimosaka2019spatiality,
  author = {Masamichi Shimosaka and Yuta Hayakawa and Kota Tsubouchi},
  title = {{Spatiality Preservable Factored Poisson Regression for Large-Scale Fine-Grained GPS-Based Population Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33011142},
}
Figure