Random Forests with a Steepend Gini-Index Split Function and Feature Coherence Injection

Abstract

Although Random Forests (RFs) are an effective and scalable ensemble machine learning approach, they are highly dependent on the discriminative ability of the available individual features. Since most data mining problems occur in the context of pre-existing data, there is little room to choose the original input features. Individual RF decision trees follow a greedy algorithm that iteratively selects the feature with the highest potential for achieving subsample purity. Common heuristics for ranking this potential include the gini-index and information gain metrics. This study seeks to improve the effectiveness of RFs through an adapted gini-index splitting function and a feature engineering technique. Using a structured framework for comparative evaluation of RFs, the study demonstrates that the effectiveness of the proposed methods is comparable with conventional gini-index based RFs. Improvements in the minimum accuracy recorded over some UCI data sets, demonstrate the potential for a hybrid set of splitting functions.