Visualizing Prediction Provenance in Regression Random Forests

Abstract

Random forest models are widely used in many application domains due to their performance and the fact that their constituent decision trees carry clear decision rules. Yet, the provenance of the predictions made by an entire forest is complex to grasp, which motivates application domain experts to adopt black-box testing strategies. We propose in this paper a coordinated multiple view system allowing to shed more light on prediction provenance, uncertainty and error in terms of bias and variance at the global model scale or at the local scale of decision paths and individual instances.