M-Tune: imbalanced data handling in machine learning by tuning the decision threshold

Overprediction of the majority class and poor predictive performance are the major issues when training machine learning classifiers with imbalanced data. This eventually leads to a larger misclassification rate and bias towards the majority class. This is a real-world problem, especially when dealing with biological datasets. There are several popular methods, including Random Under Sampling (RUS), Synthetic Minority Oversampling Technique (SMOTE), and Generalized tHreshOld ShifTing (GHOST). However, these methods either articulate the no. of samples or are biased towards the majority class. In this study, we developed a novel strategy, "M-Tune", a simple yet powerful technique for addressing the class imbalance problem. This procedure employs an ensemble approach that incorporates threshold shifting and majority voting. This new method can effectively predict the minority class. The developed method was rigorously tested with 138 drug discovery datasets. We further rigorously evaluate the performance of the M-Tune method for two classifiers with eleven different fingerprints and find that classifiers benefit from M-Tune in effectively classifying the minority class. The M-Tune method outperformed the other strategies, including the GHOST method and SMOTE and achieved comparable performance with RUS in identifying minority classes. In conclusion, we demonstrated a new threshold procedure that can be effectively applied to real-world imbalanced datasets, i.e., drug discovery data sets where the class datasets are highly imbalanced. This method aids in identifying the minority class in highly imbalanced datasets, in general, and in drug discovery datasets, in particular. Although M-Tune increases false positive predictions relative to conservative thresholding strategies, it provides a practical advantage in applications where maximizing active compound recovery is prioritized.