Spatio-Temporal Clustering and Forecasting of Urban Incident Patterns using Location-based Insights

Public safety is seriously threatened by the high rate of criminal activity. Because current prediction techniques typically rely on K-means clustering and basic classification models, evaluating large, complex crime datasets can be difficult for law enforcement agencies. The ability of these methods to produce precise, date-specific forecasts is limited because they largely ignore the dynamic, time-based behavioral patterns of various locations. To produce thorough and intelligible crime forecasts, this study suggests a novel hybrid methodology. Our method presents a two-phase system. Using a shape-based algorithm called k-Shape clustering, we first divide police patrol zones into discrete groups according to how similar their daily crime time-series "shape," as opposed to just volume. This enables us to distinguish and classify regions with distinct behavioral patterns (e.g., "stable low-level activity" versus "weekenddriven spikes"). Second, a group of specialized LightGBM (Light Gradient Boosting Machine) forecasting models can be developed thanks to this clustering, one for every recognized crime pattern. To accurately forecast future crime counts for both the overall total and ten distinct crime categories, these models are trained on a wealth of features, such as rolling averages, temporal lags, weather data, and holiday indicators. Additionally, SHAP (SHapley Additive exPlanations) is integrated into the system to offer location-based, explicable insights into the variables influencing each prediction. The final product is intended to provide law enforcement with precise, actionable intelligence for the purpose of allocating resources in a proactive manner.