A novel surrogate model called Neural-Dissected Decision Tree (NDT) to inhibit the Rashomon effect of MLPs. Our paper is currently under review by IEEE Transactions on Computational Social Systems.

Surrogate models are commonly adopted in explainable artificial intelligence (XAI). By Learning the relationships between the inputs and outputs of the black-box model, surrogate models such as decision trees can be used to inspect the intrinsic decision-making of the host model. However, different surrogate models or random initialization states may result in different explanations, which is called the Rashomon effect. For this problem, we propose a novel surrogate model called Neural-Dissected Decision Tree (NDT) to inhibit this effect of the deep-learning models.

Both decision trees are extracted from the multi-layer perceptron (MLPs). However, the decision-maker is confused when confronted with two tree explanations that are approximately equally accurate.

As Rtree is built in a strategy-level order manner which synchronized with the hidden-layers, users can better understand the decision-making of the deep-learning model through visualization.

decision tree (left) vs Rtree (right)

Please refer to the Find01_example.ipynb for details.