Most crime hotspot prediction tools rely primarily on past incident records, while environmental criminology holds that the physical character of a place shapes risk in its own right. This study introduces the Street Segment Risk Score (SSRS), which measures how strongly the features sitting on each street segment, such as pubs, bus stops, and fast-food outlets, draw risk toward it, with values that shift across police shifts and differ between weekdays and weekends. The authors tested it on robbery and theft in downtown Chicago using five learning algorithms, finding that the SSRS improved accuracy across most models, while simply counting how many such features sit on a segment, without measuring how much each one draws crime, degraded performance. The method also excludes income, unemployment, and racial composition, the variables that draw most criticism of predictive policing systems.

Research Objectives

- Introduce the Street Segment Risk Score (SSRS), a measure of risk for each street segment whose value varies with the time of day and the day of the week.

- Develop the Local Risk Score (LRS), which measures how densely crime clusters around each urban feature and weights that influence by distance along the street network.

- Test whether adding the SSRS improves prediction, both daily and by shift, against models without the score and models given only raw feature counts.

- Connect crime prediction to environmental criminology by combining the view that crime repeats where it has occurred with the view that certain places attract crime regardless of their record.

Methodology

- The study area is Chicago's Central Side, which recorded a crime rate of 466.65 per 1,000 people against a citywide average of 374.86, theft in particular.

- Records of 2,016 robberies and 31,493 thefts from the City of Chicago Data Portal between 2015 and 2018, with the 2015 data building the score and the 2016 to 2017 data training and testing the models across 2,459 street segments, analysing five types of feature: bus stops, grocery stores, pubs, fast-food restaurants, and gas stations.

- Constructing the SSRS used a bivariate network K function to screen out features with no significant effect, then segmented regression to establish how far each feature's influence reaches, producing values broken out by three police shifts and split between weekdays and weekends.

- The algorithms compared fall into three tiers: network-aware models (Graph WaveNet and STGCN), a conventional deep learning model (LSTM), and two baselines used for comparison (MLP and STNetKDE). Performance was measured with the Mean Hit Rate (MHR) at 20 percent coverage.

Key Findings

- For robbery, Graph WaveNet with the SSRS performed best on the daily models at an MHR of 58 percent, while STGCN with the SSRS led the shift models at 59.9 percent. Every algorithm gained accuracy from the SSRS in shift-level prediction.

- For daily theft, GWNet with the SSRS reached the highest MHR at 80 percent, while at shift level MLP, the simplest model in the set, outperformed every other model at nearly 80 percent. The authors attribute this to how tightly theft concentrates, with half of all incidents occurring on just 4.92 percent of the study area's street segments.

- The SSRS did not help in every case. For daily robbery it reduced MLP from an MHR of 40 percent to 23 percent, leading the authors to conclude that the score benefits most models, not all of them.

- Substituting raw feature counts for the SSRS worsened results in nearly every case, with MLP falling 53 percent on shift-level robbery. This indicates that what determines risk is not the presence of a given feature but the degree to which it draws crime at a given time.

- The graph learning algorithms doubled or tripled the performance of the traditional models, because they capture how risk spreads between street segments. The LRS analysis found pubs exert the strongest influence on robbery during the late shift and at weekends, fast-food restaurants during the day and evening shifts, while gas stations showed no significant influence.

Recommendations

- Treat the SSRS as an environmental baseline and modulate deployment according to recent incidents. Late-night patrols should concentrate on high-scoring segments near pubs and grocery stores, while daytime patrols move to segments near fast-food outlets and bus stops.

- Match prevention measures to the crime type. Segments at high risk of theft suit access control, while segments at high risk of robbery suit increased police presence.

- Train place managers, both publicans and shopkeepers, to recognise how the risk around their premises changes across the day and the week.

- Use the scores responsibly. Although the SSRS excludes biased variables, the authors warn that place-based scoring can still stigmatise an area, and recommend presenting scores as relative, provisional estimates alongside a statement of uncertainty.

- Future research should validate the SSRS in other cities alongside security experts, work at finer temporal granularity, extend the score to violent crimes, and test performance during and after COVID.

Key Takeaways

- Crime risk does not rest on incident history alone. The features present on a street shape a risk that shifts by the hour and by the day of the week.

- A theory-driven risk score improved prediction accuracy, while raw counts of those same features degraded it by as much as 53 percent.

- Excluding income, unemployment, and race answers a standing criticism of predictive policing systems. For Thailand, districts with a dense concentration of nightlife venues resemble the study area, and a street-segment score broken out by police shift would suit patrol planning better than dividing an area into a coarse grid.