Comparative Machine Learning Models for Street Crime Prediction Using Lazy Learning Algorithms: A WEKA-Based Empirical Study

Abstract

Predicting street crime levels enables proactive policing and efficient allocation of resources. Three lazy learning algorithms, namely, k-Nearest Neighbor (IBk), K-Star and Locally Weighted Learning (LWL), are compared for community-level violent crime classification (here used as a measurable proxy for street crime) within the Low, Medium, and High categories. The Communities and Crime dataset was obtained from the UCI Machine Learning Repository and contains 1,993 cleaned community instances with 100 socio-economic and law-enforcement variables. The target attribute was discretized into three nearly balanced classes using equal-frequency binning to form the "violent crimes per 100,000 population" attribute. The implementation and evaluation of all models are performed using WEKA 3.6.14 with stratified 10-fold cross-validation and using the same 10 stratified folds for each model to allow paired statistical testing. Results demonstrate that IBk with k = 7 achieves the highest accuracy at 64.53% (95% CI: 62.47–66.58%), a weighted F-measure of 0.643, and a Kappa statistic of 0.468. This improvement in comparison with K-Star and LWL is confirmed by a corrected resampled paired t-test to be statistically significant (p < 0.005 for both comparisons) with a large effect size (Cohen's d > 1.7). Overall, LWL exhibited the poorest performance and does not correctly classify any instance from the Medium class; results similar to that of its DecisionStump base learner which is limited in expressive ability. The results in this paper indicate that neighborhood averaging instance-based learners outperform kernel-weighted and entropy-based lazy learners for this multi-class task for the prediction of crime level under the experimental settings used in this paper.