Machine Learning for Crime Classification: A Fairness-Aware Approach to Class Imbalance

Automated crime classification is critical for law enforcement resource allocation. However,crime datasets exhibit severe class imbalance and fairness issues inherent in historical policing patterns. This paper presents a comprehensive machine learning framework addressing class imbalance through three strategies (adaptive class weighting, SMOTE with Borderline, SMOTE with Tomek) evaluated via nested 5 by 3-fold cross- validation on three independent datasets (CityX, San Francisco, Chicago). We compare traditional ensemble methods with deep learning baselines (LSTM, BiLSTM) and conduct rigorous fair- ness analysis. XGBoost with SMOTE and Borderline achieved best performance (94.1 percent accuracy,0.922 macro F1, statisti- cally significant with p less than 0.001) with 8.7 percent minority class improvement and 3.2 times faster inference than LSTM. Critical fairnessanalysis revealed demographic parity disparities (10.3 percent gap) requiring human oversight. Feature ablations demonstrate top-10 features sufficient; spatiotemporal patterns analyzed. Results show strong generalization across datasets (mean accuracy 93.4 percent plus or minus 1.8 percent). Com- putational analysis reveals XGBoost achieves optimal accuracy- efficiency tradeoff. This work provides rigorous evidence for responsible AI deployment in crime classification with explicit fairness considerations