Pareto optimization of textual and engagement features for crime-related social media analytics

Crime-related social media content provides valuable information for situational awareness and policy support; however, engagement-driven signals, such as likes and retweets, can introduce popularity bias and affect predictive outcomes. This study proposes a fairness-aware, Pareto-based, multi-objective optimization framework for feature selection and Random Forest hyperparameter optimization, leveraging textual and engagement features. A dataset of 32,000 crime-related tweets covering Fraud, Accident, and Murder was represented using TF-IDF and interaction-based features. Four optimization strategies, Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), Evolutionary Reinforcement Learning (ERL), and Non-Dominated Sorting Genetic Algorithm II (NSGA-II), were comparatively evaluated. NSGA-II jointly maximized classification performance while minimizing a fairness penalty based on Statistical Parity Difference (SPD), Disparate Impact (DI), and Equality of Opportunity Difference (EOD). In this study, fairness refers to mitigating engagement-driven popularity bias across behavioral user groups rather than demographic discrimination. NSGA-II achieved an accuracy of 0.8118 and an F1-score of 0.8051, with an SPD of 0.08, DI of 0.97, and EOD of 0.06, demonstrating the most favorable balance between predictive performance and behavioral fairness. Although removing engagement features improved standalone Accuracy and F1-score, retaining them enabled explicit assessment and mitigation of engagement-driven bias. Stratified cross-validation and comparative analysis across the optimization strategies further supported the consistency of the observed results. The findings demonstrate that integrating fairness objectives directly into Pareto-based optimization provides a practical and interpretable approach to balancing predictive effectiveness and engagement-driven fairness in crime-related social media analytics.