Reproducibility, the ability to reproduce computational results using identical data and software, is a cornerstone of the scientific methodology. However, through the past decade, several studies revealed a widespread lack of results’ reproducibility, to the point that the existence of a reproducibility crisis is now acknowledged in various fields. In Machine Learning, given the flexibility available in various phases of constructing a computational model, the experiments are not immune to reproducibility issues either. In case of imbalance learning for problems with multiple classes, the problem is even more severe since there are more parameters in play for constructing a model. The resulting reproducibility challenges have implications in various disciplines including bioinformatics, the primary focus of our study.

Researchers have already taken counter-measures proposing various recommendations for having results’ reproducibility in this domain of study. Some conferences have even adopted new measures in that regard. Following those guidelines could ensure reproducibility to an agreeable degree in balanced problems. In this work we demonstrate that in an imbalanced scenario, even in its basic form, a study report with a fair amount of details, could reproduce a wide range of results if methodological flexibility is permitted.