Performance Assessment and Interpretation of Random Forests by Three-dimensional Visualizations

Abstract

Ensemble learning techniques and in particular Random Forests have been one of the most successful machine learning approaches of the last decade. Despite their success, there exist barely suitable visualizations of Random Forests, which allow a fast and accurate understanding of how well they perform a certain task and what leads to this performance. This paper proposes an exemplar-driven visualization illustrating the most important key concepts of a Random Forest classifier, namely strength and correlation of the individual trees as well as strength of the whole forest. A visual inspection of the results enables not only an easy performance evaluation but also provides further insights why this performance was achieved and how parameters of the underlying Random Forest should be changed in order to further improve the performance. Although the paper focuses on Random Forests for classification tasks, the developed framework is by no means limited to that and can be easily applied to other tree-based ensemble learning methods.