Robust Thresholding Strategies for Highly Imbalanced and Noisy Data

Many studies have shown that non-default decision thresholds are required to maximize classification performance on highly imbalanced data sets. Thresholding strategies include using a threshold equal to the prior probability of the positive class or identifying an optimal threshold on training data. It is not clear, however, how these thresholding strategies will generalize to imbalanced data sets that contain class label noise. When class noise is present, the positive class prior is influenced by the class label noise, and a threshold that is optimized on noisy training data may not generalize to test data. We employ four thresholding strategies: two thresholds that are optimized on training data and two thresholds that depend on the positive class prior. Threshold strategies are evaluated on a range of noise levels and noise distributions using the Random Forest, Multilayer Perceptron, and XGBoost learners. While all four thresholding strategies significantly outperform the default threshold with respect to the Geometric Mean (G-Mean), three of the four thresholds yield unstable true positive rates (TPR) and true negative rates (TNR) in the presence of class noise. Results show that setting the threshold equal to the prior probability of the noisy positive class consistently performs best according to G-Mean, TPR, and TNR. This is the first evaluation of thresholding strategies for imbalanced and noisy data, to the best of our knowledge, and our results contradict related works that have suggested optimizing thresholds on training data as the best approach.