ENHANCING AI SYSTEM TRANSPARENCY AND EXPLAINABILITY: INTEGRATING FORMAL METHODOLOGIES FOR IMPROVED MODEL PERFORMANCE AND INTERPRETABILITY

Abstract

Artificial Intelligence operates as essential business infrastructure in healthcare together with finance and autonomous systems. The output decisions from deep learning neural network-based AI models present significant barriers to both understanding and interpretation. The absence of explainability features between models creates trust-related conflicts for users and regulators and directly affected industrial stakeholders. The combination of SHAP and LIME presents viable explanation tools but produces imprecise interpretations when evaluated against high-dimensional real-time datasets. Random Forest surpassed both Logistic Regression and SVM by obtaining superior results in generalization testing which produced greater training and validation accuracy levels. The accuracy measurements revealed that Random Forest achieved 0.894 training accuracy along with 0.879 validation accuracy while Logistic Regression maintained 0.905 training accuracy and 0.874 validation accuracy and SVM achieved 0.848 training accuracy with 0.867 validation accuracy. The decision outcomes from the model were primarily influenced by Features 3 and 6 according to SHAP and LIME analysis. Random Forest presented the best ROC and precision-recall curves which indicated its strength to separate distinct classes. Future research should optimize the methodologies through development that enables their scaling across multiple applications while achieving better performance specifically in time-sensitive and dimensionally complex systems. Despite these promising results, the study encountered two primary limitations: Formal methods face scalability issues and all models displayed poor AUC scores as their primary limitations. Both Logistic Regression and Random Forest with SVM yielded prediction performance similar to random guessing based on AUC scores of 0.51 and 0.50 respectively. The research focus should optimize scalable methods aimed at improving performance while solving time-sensitive high-dimension problems.