The study of Rashomon effects on machine learning : a case study on breast cancer

Abstract

The Rashomon effect is a theory that suggests the presence of multiple uncorrelated observations and explanations that can be made for a single observation. This theory has been translated into a popular machine learning method: Random Forests which uses bootstrapping (bagging) algorithms to create a set of uncorrelated decision trees that together make the decision (prediction) of the final result. In this study, we will be using 3 ER breast cancer datasets as a case study and we look at the results of the selection of each individual tree in the forest using the standard random forest algorithms and when bootstrapping of the attributes was removed. We found that most forests converged into a few highly correlate gene signatures which dominates the prediction and masks the errors of non-accurate models. Besides, because the random forest algorithm can generate highly accurate with a group of and non-predictive signatures, we need to be careful when using random forest machine models for prediction in the field of cancer biology.