Predicting Urban Crime with Minimal Spatiotemporal Features Using Machine Learning

This study evaluates the efficiency of using borough and time to predict long-term crime patterns in London. The data was taken from Metropolitan Police Service records from 2008-2016. Furthermore the crime counts were aggregated annually at borough level and modeled as a function of borough and year. The aim of the study was to compare different classical machine learning, neural networks and ensemble approaches, all while testing the stability of spatial crime patterns. Linear regression, Ridge regression, Decision Tree, Random Forest, K-Nearest Neighbors, a Multilayer Perceptron, and a TensorFlow neural network were evaluated using MSE, MAE, and MAPE, alongside a mean-value baseline predictor. The results suggest that nonlinear models, particularly Decision Trees and ensemble methods out perform the baseline. This indicates structural differences between boroughs and gradual temporal change. Essentially this indicates that borough-level crime variation can be explained by minimal spatiotemporal information alone. This illustrates the predictive capacity of simple spatial-temporal structure. Furthermore the ethical need to interpret forecasts as recorded crime patterns instead of inherent characteristics of each borough.