Machine learning for crime classification: A fairness-aware approach to class imbalance

Automated crime classification is critical for law enforcement resource allocation, yet crime datasets exhibit severe class imbalance and fairness issues rooted in historical policing patterns. This paper presents a methodologically rigorous machine learning framework addressing these challenges through three strategies: adaptive class weighting, SMOTE-NC (for mixed categorical/numerical tabular data), and SMOTE-NC with Tomek link removal. Models are evaluated via nested spatiotemporal cross-validation-spatial block outer loop combined with forward-chaining inner loop-on three independent datasets: the Rajshahi Metropolitan Police (RMP) incident dataset, San Francisco, and Chicago. We compare tabular-native architectures (XGBoost, CatBoost, TabNet) with in-processing fairness baselines (FairXGBoost, FairGBM). Protected demographic attributes are strictly excluded from training features and reserved solely for post-hoc fairness auditing. FairXGBoost with SMOTE-NC achieved the best fairness-accuracy tradeoff (92.4% accuracy, 0.901 macro F1, p < 0.001) with an 8.1% minority class recall improvement and a demographic parity gap reduced to 5.8%. We explicitly discuss the Impossibility Theorem of Fairness, sociological grounding of dominant predictive features, the human cost of false positives, the dark figure of crime, and alignment with EU AI Act and UCR/NIBRS taxonomy standards. Multidataset results demonstrate strong generalization (mean accuracy 93.1% ± 1.6%) while maintaining

responsible deployment standards.