An Adaptive Sampling Framework for Robust Anomaly Detection in Overlapping and Imbalanced Datasets

초록

This study introduces an adaptive sampling framework for robust anomaly detection in overlapping and imbalanced datasets. The proposed approach integrates geometric oversampling with ensemble-based undersampling to overcome the limitations of conventional resampling techniques. Unlike interpolation-based methods such as SMOTE, the geometric oversampling strategy preserves the local structure of minority class data and minimises the generation of ambiguous synthetic samples. In parallel, the ensemble-based undersampling mechanism selectively retains majority instances to balance recall and precision, leading to consistent improvements in the F1-score across classifiers. Feature selection is incorporated to enhance model efficiency and interpretability by eliminating redundant variables and emphasising informative predictors. Experiments on the UCI Default of Credit Card Clients dataset, along with additional validation on the UCI Bank Marketing dataset, demonstrate that the framework achieves superior performance and generalisability under severe class imbalance and overlapping boundaries. From an economic cybernetics perspective, the proposed method strengthens decision reliability in high-stakes domains such as credit risk assessment, fraud detection, and cybersecurity.

키워드