Robust Anomaly Detection under Severe Class Imbalance: An Adaptive Geometric Sampling Approach

Overview

Abstract
Detecting rare but high-impact anomalies in financial data remains a fundamental challenge due to severe class imbalance and overlapping class distributions. Conventional resampling techniques such as SMOTE often improve recall at the expense of precision by generating ambiguous synthetic samples near decision boundaries. In this talk, I present an adaptive sampling framework that integrates geometry-aware oversampling with ensemble-based undersampling to address these limitations.
 

The proposed geometric oversampling method generates synthetic minority samples within localized regions defined by the intrinsic structure of abnormal data, thereby preserving class topology and reducing boundary noise. This is combined with an ensemble undersampling strategy that selectively retains informative majority instances, enabling a stable trade-off between recall and precision. Feature selection is further incorporated to enhance efficiency and interpretability. Empirical results on a benchmark credit default dataset demonstrate consistent improvements in recall and F1-score across classifiers, highlighting the effectiveness of the proposed framework in practical financial anomaly detection problems.