A computational criminology study published in Crime Science introduces GraphVenn, an open-source algorithm designed to solve the long-standing challenge of identifying the globally optimal placement of fixed-radius crime hotspots. Criminological research has repeatedly established that unlawful acts concentrate disproportionately within a tiny fraction of micro-places, a phenomenon known across the discipline as the law of crime concentration. Modern hotspot policing, strategic surveillance placement, and situational prevention initiatives depend heavily on identifying these critical geographic nodes. However, until now, researchers and municipal law enforcement agencies lacked computational tools capable of determining whether empirical hotspot selections reached the absolute upper bound of crime coverage theoretically possible. GraphVenn addresses this exact problem by modeling empirical crime patterns directly through spatial graph theory and exact mathematical programming, providing researchers with both a benchmark for theoretical maximum coverage and a practical tool for operational deployment.
Traditional spatial hotspot identification has relied primarily on approximate techniques, including kernel density estimation, greedy ranking heuristics, and spatial scan statistics. Although these conventional methods remain widely utilized across municipal police agencies, they introduce substantial technical limitations. Many established techniques impose arbitrary grid discretization across urban landscapes, artificially bisecting high-density crime clusters between neighboring cells and systematically underestimating their true localized intensity. Furthermore, greedy ranking frameworks evaluate candidate locations iteratively one at a time, failing to explore the vast combinatorial solution space required to optimize multiple hotspots simultaneously. Conventional tools cannot guarantee that a selected portfolio of hotspots represents the true global optimum. Consequently, crime analysts have operated without knowing how many reportable incidents were overlooked due to sub-optimal hotspot positioning, heuristic approximations, or rigid spatial boundaries that fail to conform to the true geometry of urban offending.
To formalize the optimization challenge, the study establishes precise mathematical formulations governing crime distributions and spatial parameters. The geographic domain contains a set of recorded crime events and a corresponding set of unique geographic coordinate locations. Because multiple incidents repeatedly occur at identical geographic coordinates, the total incident count substantially exceeds the number of unique locations. The algorithm requires users to define three fundamental parameters: the hotspot coverage radius in meters, the target number of hotspot centers to return, and the spatial resolution parameter denoting coordinate decimal precision. For example, a spatial resolution setting of three decimals corresponds to approximately one hundred eleven meters between sampling points, four decimals yields roughly eleven meters, and five decimals establishes fine-grained spacing of approximately one meter. The analytical objective centers on identifying the exact spatial points whose circular boundaries maximize the total count of distinct crime events captured without duplicate counting.
GraphVenn executes its spatial detection task across three sequential computational phases. In the first phase, the algorithm constructs an undirected, weighted graph over the complete set of unique crime coordinates. Nodes represent individual physical locations annotated with historical incident counts, while edges link pairs of locations separated by a distance less than or equal to the designated radius, weighted by exact physical distance. This structure naturally restricts the relevant geometric search space to the circular neighborhood surrounding empirical incident locations, eliminating barren urban space where no crimes have occurred. Additionally, the formulation leverages the mathematical principle of two-radius disjointness. When two distinct candidate locations are separated by more than twice the specified radius, their coverage zones cannot intersect. This mathematical property allows the algorithm to partition sprawling citywide crime patterns into independent, disconnected components that can be solved separately without risking conflicting incident assignments.
The second phase establishes candidate sampling locations using a technique designated as Venn voting. Rather than scanning entire municipal territories across continuous coordinates, GraphVenn restricts its sampling grid exclusively to the union of radius-bounded zones encircling known crime points. Within each relevant zone, candidate evaluation points are spaced according to the selected coordinate resolution. Each candidate point tallies a cumulative vote calculated from the aggregate crime counts of all known incident locations falling within its surrounding circular radius. As intersecting circular neighborhoods overlap across dense urban sectors, these votes accumulate rapidly. This mechanism allows the algorithm to systematically discover intermediate geographic positions situated between historical crime clusters. By shifting hotspot centers toward these calculated intermediate coordinates, the system frequently captures multiple adjacent incident clusters simultaneously, yielding dramatic empirical increases in distinct incident coverage compared to methods centered strictly on isolated historical coordinates.
In the third phase, GraphVenn engages its core optimization engine to select the top hotspot locations from the sampled candidates. The algorithm provides two distinct operating modes: a globally optimal strategy and a fast greedy approximation. In optimal mode, candidates undergo rigorous pruning via mathematical dominance tests and upper-bound eliminations before being partitioned into independent connected components. GraphVenn formulates the selection within each component as a maximum coverage problem solved exactly through integer linear programming using the open-source Coin-OR Branch-and-Cut solver. A dynamic programming routine then distributes the allocated hotspot budget across components, guaranteeing that the final output captures the absolute maximum number of distinct incidents possible under the chosen parameters. Conversely, the greedy variant utilizes a lazy-greedy routine that iteratively chooses candidate centers offering the highest marginal gain in distinct crime incidents, bypassing linear programming solvers entirely.
The research team evaluated GraphVenn using three consecutive years of historical crime incident records drawn from three distinct municipal environments: Malmö, Sweden, alongside Boston and New York City in the United States. In total, the experimental evaluation examined 1,746,917 reported crime incidents. The Malmö dataset comprised 72,443 incidents spanning 2018 through 2020 at full spatial precision, ethically approved by the Swedish Ethical Review Authority. The Boston dataset included 221,677 incidents from 2019 through 2021, while New York City contributed 1,452,797 events spanning 2016 through 2018, both obtained from the Crime Open Database where coordinates are aggregated to street intersections. Computational experiments were executed on a consumer-grade laptop equipped with an Apple M2 Max processor and thirty-two gigabytes of memory. GraphVenn was compared directly against three established baselines: kernel density estimation, greedy Predicted Accuracy Index maximization, and the graph-based GraphTrace algorithm.
Evaluation of computational feasibility demonstrated that global optimality is practically achievable at city scale without supercomputing infrastructure. Testing a standard one-hundred-meter radius and targeting one hundred hotspots, optimal GraphVenn at fine-grained one-meter resolution completed annual analyses in roughly two and a half hours for Malmö and Boston, and just under five and a half hours for New York City. When spatial resolution was adjusted to four decimals, optimal runtimes dropped dramatically to under thirteen minutes for Malmö, eight minutes for Boston, and sixteen minutes for New York City. The fast greedy mode delivered even more striking speedups, processing four-decimal configurations across all three cities in under one minute. Memory consumption remained well within ordinary desktop parameters, ranging from approximately 4.1 gigabytes at coarse two-decimal resolution up to 15.2 gigabytes for New York City under fine one-meter sampling.
Comparative coverage evaluations revealed that GraphVenn substantially outperformed all traditional baseline mapping approaches across every evaluated city. In optimal mode, the algorithm established the true mathematical ceiling for distinct incident coverage. In New York City, optimal GraphVenn captured 51,522 distinct crimes within its top one hundred hotspots, whereas kernel density estimation captured only 35,098 crimes and GraphTrace captured 28,241 crimes. The greedy Predicted Accuracy Index baseline had to be terminated entirely for New York City because its projected runtime exceeded two weeks. In practical operational terms, the traditional baselines missed between 16,000 and 23,000 crimes that were fully capable of being encompassed within identical surface areas. Across all three cities, the greedy GraphVenn variant reached within 0.1 to 1.9 percent of the absolute global optimum while running up to two orders of magnitude faster than the exact integer programming solver.
Detailed ranking analyses in Malmö and New York City illustrated fundamental differences in how global optimization distributes spatial resources compared to heuristic techniques. In top-tier rankings, greedy heuristics and competing algorithms occasionally recorded higher individual incident totals within their single highest-ranked hotspot. However, this occurred because heuristic methods greedily consume dense incident clusters without considering subsequent spatial overlap. In contrast, GraphVenn’s exact integer linear programming engine deliberately sets aside certain high-density candidates when selecting them would create overlapping coverage, thereby preserving incident availability for subsequent hotspots. By redistributing selections across lower-ranked positions, the optimal approach sustained substantially higher crime totals across ranks ninety-one through one hundred. This mathematical coordination ensured that the complete portfolio maximized total distinct crime coverage across the entire jurisdiction, avoiding the premature exhaustion of overlapping incidents common to greedy selection.
The authors outline several methodological boundaries alongside promising pathways for future research. The current algorithm restricts coverage geometries to fixed-radius circular footprints, omitting irregular polygons, linear street networks, or natural urban barriers such as rivers, major highways, and architectural structures. Furthermore, because municipal crime geocoding is often spatially aggregated to intersections, sampling resolutions finer than eleven meters provide limited practical utility while inflating computational overhead. The researchers suggest extending future iterations by merging adjacent evaluation points into contiguous, irregular hotspot zones, dynamically adjusting coordinate resolution to match local incident density, and incorporating jurisdictional administrative boundaries such as police beats. Because the underlying standalone impact scores reflect stable local crime intensities, they could also serve as high-resolution inputs or priors for spatio-temporal forecasting models, bridging descriptive spatial optimization and predictive policing strategies.
Subject of Research: Globally optimal spatial algorithm for detecting and ranking fixed-radius crime hotspots
Article Title: A globally optimal algorithm for hotspot detection and ranking
Article References: Boldt, M. (2026). A globally optimal algorithm for hotspot detection and ranking. Crime Science, 15(1), Article 7. https://doi.org/10.1186/s40163-026-00269-x
Image Credits: AI Generated
DOI: 10.1186/s40163-026-00269-x
Keywords: crime science, hotspot policing, spatial optimization, graph algorithms, integer linear programming, computational criminology, urban crime mapping, geographic information systems, globally, optimal, algorithm, hotspot
Cite Scienmag News
Courtney Benton. (November 9, 2026). Novel Optimization Algorithm Identifies Globally Optimal Crime Hotspots at City Scale. Scienmag. https://scienmag.com/novel-optimization-algorithm-identifies-globally-optimal-crime-hotspots-at-city-scale/
Courtney Benton. "Novel Optimization Algorithm Identifies Globally Optimal Crime Hotspots at City Scale." Scienmag, 9 November 2026, https://scienmag.com/novel-optimization-algorithm-identifies-globally-optimal-crime-hotspots-at-city-scale/. Accessed 17 November 2026.
Courtney Benton. "Novel Optimization Algorithm Identifies Globally Optimal Crime Hotspots at City Scale." Scienmag. November 9, 2026. https://scienmag.com/novel-optimization-algorithm-identifies-globally-optimal-crime-hotspots-at-city-scale/








