Instils Trust in Random Forest Predictions
Gopal Singh Jamnal · 2023
This paper addresses the interpretability and transparency behind the random forest model predictions. Random forest is an ensemble of bootstrapped independent decision trees that are trained on subsets of input data to make predictions. Although random forest is a robust model that can overcome bias, its inherent complexity, and poor interpretability can make it challenging to apply in many application domains that require transparency and explainability in the model’s predictions. This lack of transparency in the decision-making process can prevent users from analyzing what makes the model arrive at a specific prediction. The paper presents a visual analytic application to overcome transparency challenges by providing a clear structure of individual decision trees and hierarchical relationships between features. This allows users to analyze latent information and instils trust in the random forest model. Additionally, statistical analysis of feature ranking agreement and prediction popularity reduces mental burden of the user. The paper includes two case studies to evaluate model’s uncertainty, bias, and variances in predictions with explainability at local and global scales of decision paths. The visual analytics application provides a coordinated multiple-view system to instil trust in random forest models.