Interpretable Predictions for Crime Categories Using Log Loss Approach for Imbalanced Target Feature
A Manasa, Snigdha Sen · 2024
Crime indicates activities that infringe laws and regulations constituted by a governing authority. These wrongdoings can range from minor aberrations such as petty theft to more serious crimes like assault, robbery, or murder. Understanding crime patterns, causes, and trends is important for creating effective law enforcement strategies and enforcing policies that facilitate to a risk-free and stable society. This paper explores the interpretability of crime category predictions generated by XGBoost and Random Forest models on San Francisco crime data ranging 12 years. Notably, in comparison to other models, such as k-Nearest Neighbors (KNN), Stochastic Gradient Descent(SGD) and Naïve Bayes, our XGBoost and Random Forest models featured superior predictive excellence, Obtaining competitive log-loss scores of 2.277761 and 2.283956, respectively, to augment interpretability we have employed Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) to clarify the decision-making processes of these models. This interpretability analysis would definitely add a deeper insight into the contributing factors of model's predictive behavior towards crime rate prediction. Additionally, we have dealt with imbalanced target feature as well. In comparison to existing research in the field, our work stands out by introducing advanced interpretability techniques, LIME and SHAP, enhancing transparency in crime prediction models. This innovative approach surpasses traditional methods, providing a more comprehensive understanding of crime dynamics and achieving superior predictive excellence. Innovation is introduced through the incorporation of advanced interpretability techniques, specifically Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP).