On the evaluation of classification quality - the robustness of the AUC of Balance Accuracy Curve (BAC) to anomalies in the classification process
Aleksandra Weiss, Marcin Młyński, Piotr Artiemjew · Procedia Computer Science · 2024
In this study, we extend our exploration of the AUC of Balance Accuracy Curve (BAC), a novel parameter we have developed that rivals the traditional metrics used for classification model evaluation, such as AUC of ROC and PR curves. BAC stands out for its straightforwardness, particularly in evaluating the overall efficacy of training systems amid varying degrees of class imbalance. Our current focus investigates the resilience of BAC against anomalies during the classification process, examining its behavior across multiple anomaly levels. For easier understanding, as a benchmark classifier we have employed the k-Nearest Neighbours (kNN) method, where we applied the most common distance metrics. The overall verification was conducted using real-world datasets selected from the UCI Repository, providing a practical context for our research.