A New Class-Weighting Formulation for the Class Imbalance Problem: A Methodological Research

Objective: Many of the machine learning classification algorithms are not robust against unbalanced classes and result in poorly accurate and biased models. One way to address class imbalance is to assign weights to classes. This article proposes a new class-weighting approach to improve the classification problem when there is an imbalance between two class. Material and Methods: The performances of the new formulation were compared with the previously proposed Inverse of Square Root of Number of Samples, effective number of samples weighting formula and unweighted Random Forest solutions. A simulation study was performed using performances of 3 imbalance rates (0.10, 0.20, 0.30), 6 different sample sizes (250, 300, 350, 400, 450, 500) and 4 different methods with 1,000 repetitions. Additionally, the methods were analyzed on the lung cancer dataset with 39 samples in the minority group and with 270 samples in the majority group. Results: Experimental results demonstrated that our proposed weighting formula, least number of ratio and range multiplier, performed equal to or better solution than Inverse of Square Root of Number of Samples in both simulations and real data. Generally, minority class accuracy and balanced accuracy of our formulation were either very close to or higher than that of Inverse of Square Root of Number of Samples. Conclusion: The new formulation provided accuracy estimates of the 2 classes in a balanced way for each sample size and for each imbalance rate. Additionally, as the sample size increased from 250 to 500, stable decreasing weights could be obtained for the patient and control groups.