AI-Enhanced Decision Intelligence: A Data-Centric Approach to Risk Prediction with Imbalanced Data

This preliminary research preprint presents an AI-Enhanced Decision Intelligence framework for cybersecurity risk prediction under imbalanced data. The study investigates how data-centric imbalance-handling strategies influence machine-learning prediction and downstream risk-oriented decision making using the UNSW-NB15 benchmark dataset. The proposed framework compares XGBoost under original-data, class-weighting, random-undersampling, and SMOTENC configurations, with Logistic Regression and Random Forest used as baseline models. Precision-Recall Area Under the Curve (PR-AUC) is designated as the primary predictive metric, complemented by recall, precision, F1-score, ROC-AUC, balanced accuracy, and Brier score. A validation-based thresholding framework is designed to translate predicted probabilities into low-, medium-, and high-risk categories and evaluate decision utility. The research implementation has been established and validated locally, including dataset acquisition, schema verification, Python dependency verification, experiment-code validation, and Git-based version control. The final cloud-based experiment and quantitative results remain pending in this preliminary version. Future work will execute the experiment, report empirical results, analyze the effect of imbalance-handling strategies, and update the manuscript accordingly.