Unsupervised Domain Adaptation for Crime Risk Prediction Across Cities

Binbin Zhou, Longbiao Chen, Sha Zhao, Shijian Li, Zengwei Zheng, Gang PanIEEE Trans. Comput. Soc. Syst 2023 · DOI

Introduction

Predicting crime risk is a cornerstone of urban safety and resource allocation. However, the effectiveness of predictive models is often hampered by a lack of labeled historical data in many cities due to privacy concerns, high collection costs, or regulatory hurdles. While transfer learning offers a potential solution by leveraging data from a "source" city to predict crime in a "target" city, it faces the significant challenge of context inconsistency. For instance, one city may provide detailed mobility data (e.g., taxi trips), while another only provides meteorological data. Standard approaches often discard these city-specific features, losing vital predictive information. This paper introduces UDAC (Unsupervised Domain Adaptation for Crime), a framework designed to bridge this gap by "hallucinating" missing local data and learning domain-invariant features.

Methodology

The UDAC framework operates through three primary phases to overcome data scarcity and context inconsistency:

  1. Inter-city Similar-Grid Matching: Cities are partitioned into equal-sized grids. The model identifies "partner" grids in the source city that share spatio-temporal similarities with target city grids based on common context data, such as weather and Point-of-Interest (POI) distributions.
  2. Auxiliary Feature Construction: To address missing data types in the target city, UDAC constructs auxiliary features. It maps city-specific data available only in the source city (e.g., taxi mobility) onto the target city's grids using the similarities established in the matching phase.
  3. Dense Convolutional Network with UDA: The core architecture uses a dense convolutional network to learn high-level representations. It optimizes three objectives: a prediction loss on source data, an adversarial domain classification loss to ensure features are domain-invariant, and a Maximum Mean Discrepancy (MMD) loss to align feature distributions between cities.
Overview of the UDAC framework, illustrating inter-city grid matching and auxiliary feature construction.

Experimental Setup and Datasets

The study evaluated UDAC using real-world 2015 data from New York City (NYC), Chicago, and Los Angeles (LA). These datasets included crime incidents, meteorological data, POI distributions, and police station locations. NYC specifically provided taxi trip data, which served as the auxiliary context for transfer tasks to cities lacking such mobility records.

Key Results

UDAC was tested across six city-pair transfer scenarios (e.g., NYC → Chicago, LA → NYC). The model consistently outperformed state-of-the-art baselines, including DANN and STCNet. In the NYC → Chicago experiment, UDAC achieved a Mean Absolute Percentage Error (MAPE) of 0.121, a substantial improvement over the standard CNN baseline of 0.448.

Furthermore, UDAC demonstrated superior ranking performance (HR@K), effectively identifying the most dangerous hotspots. The model is also computationally efficient, with an average execution time per test sample of less than 0.5 seconds, making it suitable for real-time urban monitoring.

Comparison of UDAC performance against the CNN baseline across various city-pair transfer tasks.

Conclusion

UDAC provides a robust framework for crime risk prediction in data-scarce environments. By reconstructing missing auxiliary features rather than discarding them, the model preserves the predictive power of urban context. Future work will explore scenarios with even higher levels of context inconsistency and aim for finer-grained predictions at the individual road segment level.

Summary updated

BibTeX
@article{zhou2023unsupervised,
  author = {Binbin Zhou and Longbiao Chen and Sha Zhao and Shijian Li and Zengwei Zheng and Gang Pan},
  title = {{Unsupervised Domain Adaptation for Crime Risk Prediction Across Cities}},
  journal = {IEEE Trans. Comput. Soc. Syst.},
  year = {2023},
  doi = {10.1109/tcss.2022.3207987},
}
Figure