Predictive Hotspot Mapping for Data-driven Crime Prediction

Predictive hotspot mapping is an important problem in crime prediction and control. An accurate hotspot mapping helps in appropriately targeting the available resources to manage crime in cities. With an aim to make data-driven decisions and automate policing and patrolling operations, police departments across the world are moving towards predictive approaches relying on historical data. In this paper, we create a non-parametric model using a spatio-temporal kernel density formulation for the purpose of crime prediction based on historical data. The proposed approach is also able to incorporate expert inputs coming from humans through alternate sources. The approach has been extensively evaluated in a real-world setting by collaborating with the Delhi police department to make crime predictions that would help in effective assignment of patrol vehicles to control street crime. The results obtained in the paper are promising and can be easily applied in other settings. We release the algorithm and the dataset (masked) used in our study to support future research that will be useful in achieving further improvements.

💡 Research Summary

The paper addresses the practical problem of generating predictive crime‑hotspot maps to support police patrolling and resource allocation. Recognizing that many police departments worldwide are moving toward data‑driven, predictive policing, the authors note a gap in existing approaches: most models either ignore temporal dynamics or lack a mechanism to incorporate expert knowledge from field officers or other real‑time sources. To fill this gap, they develop a non‑parametric spatio‑temporal kernel density estimation (KDE) framework that simultaneously models spatial location, intra‑day timing, and inter‑day trends, while allowing expert inputs to be blended in a Bayesian manner.

Data and Setting
The study uses Police Control Room (PCR) call records for street crimes in Delhi from October 2019 to March 2021. Each record contains a timestamp and a geocoded location. Although the raw database contains millions of potential grid points, the authors focus on the few hundred actual incident points per day, which makes the KDE computationally tractable. The dataset was cleaned to remove prank calls and to standardize location fields; it is not publicly available, but a masked version has been released on Kaggle for reproducibility.

Model Construction

1. Spatial Kernel – A two‑dimensional Gaussian kernel is applied, with an adaptive bandwidth that shrinks in dense clusters and expands in sparse areas, preventing over‑fitting while preserving local detail.
2. Intra‑day Timing – A circular kernel (e.g., von Mises) captures the periodic nature of time‑of‑day, ensuring continuity between late‑night and early‑morning hours.
3. Inter‑day Trend – The authors introduce block‑weighted temporal smoothing: recent weeks are grouped into blocks (1‑week, 2‑week, 4‑week) and assigned weights that decay with age. These weights are treated as Bayesian priors and updated from the data, allowing the model to balance short‑term spikes against longer‑term patterns.
4. Expert Input Integration – Field officers can provide “risk increase” or “risk decrease” annotations for specific locations and time windows. These annotations modify the prior weights of the corresponding blocks, effectively shifting the posterior density toward or away from expert‑identified hotspots.

Parameter estimation is performed via a customized Expectation‑Maximization scheme that handles the non‑linear combination of spatial, circular, and block‑weighted components. Computational efficiency is achieved by limiting KDE calculations to observed event points; the algorithm scales linearly with the number of incidents rather than the total number of grid cells, making weekly updates feasible on standard hardware.

Experimental Evaluation
Four model variants are compared: (a) pure spatial KDE, (b) spatial + circular timing, (c) spatial + block weighting, and (d) the full integrated model. Performance is measured using precision, recall, and F1‑score on a held‑out week‑by‑week forecast. The full model consistently outperforms the baselines, achieving 8–12 percentage‑point gains in F1‑score across all weeks. The most pronounced improvements appear during the high‑crime evening window (20:00–24:00), where temporal dynamics are strongest. Simulated expert inputs (e.g., marking a construction site as high‑risk) further boost performance by roughly 4 % points, demonstrating the value of the Bayesian expert‑integration mechanism.

Insights for Policing

• Joint modeling of location and intra‑day timing yields better predictions than a purely spatial approach, confirming that crimes at a given hour are influenced by patterns at other hours.
• While Delhi police’s routine patrol allocations are relatively stable week‑to‑week, the model reveals that hotspot locations can shift noticeably each week, suggesting a need for more dynamic scheduling.
• Traditional “key” locations (metro stations, temples, markets) are indeed frequent hotspots, yet the model uncovers additional vulnerable spots that are not currently prioritized, offering actionable intelligence for reallocating patrol vehicles.
• Expert inputs, even when simulated, improve predictive accuracy, validating the collaborative framework that respects officer expertise while leveraging historical data.

Limitations and Future Work
The authors acknowledge that KDE treats space as homogeneous, ignoring land‑use, road networks, barriers, or natural features that can affect crime dispersion. Moreover, the model does not capture inter‑event dependencies such as retaliatory crimes. To address these issues, future research will explore incorporating spatial weight matrices, integrating point‑process models like log‑Gaussian Cox processes or Hawkes processes, and blending them with the adaptive KDE to capture both intensity and interaction effects. Validation of expert inputs through systematic collection and bias assessment is also planned.

Conclusion
The paper delivers a practical, computationally efficient, and empirically validated framework for spatio‑temporal crime hotspot prediction that blends data‑driven density estimation with expert knowledge. Its successful deployment with the Delhi police demonstrates that such a system can enhance patrol planning, improve public safety, and serve as a template for other smart‑city policing initiatives worldwide.