A Text-Semantic Stacked Ensemble Framework for Bias-Reduced Defence Clearance Decision Support

Abstract

The exponential growth of digital records in legal and defence sectors has created an urgent demand for intelligent systems capable of extracting insights from large-scale unstructured data. Security clearance adjudications involve complex case digests, categorical records, and appeals that require consistent, unbiased, and accurate decision support. Traditional manual methods are time-intensive and often subject to inconsistencies, while conventional machine learning models struggle with imbalanced datasets and fail to capture contextual nuances in natural language. This research addresses these challenges by integrating natural language processing (NLP) with data balancing techniques and ensemble classification strategies to enhance predictive reliability in defence clearance adjudications. The study focuses on two binary classification tasks: Favorable Decision (Yes/No, indicating whether the appeal's decision grants or rejects clearance) and Decision Upheld (Yes/No, indicating whether the original decision is sustained or overturned). Existing baseline models such as K-Nearest Neighbors (KNN), Logistic Regression (LR), and Multinomial Naïve Bayes (MNB) were evaluated, but these individual classifiers were limited in handling imbalanced data and contextual variations in case digests. To overcome these drawbacks, the proposed methodology integrates NLP-driven preprocessing with the Synthetic Minority Oversampling Technique (SMOTE) and employs a Stacked Classifier, where Random Forest (RF) serves as the base learner and Logistic Regression (LR) acts as the meta-classifier. This stacking framework effectively leverages the feature learning capability of RF and the decision boundary optimization of LR, resulting in improved predictive accuracy and balanced classification across both Yes/No target classes.