back to list

Project: Visualization of Oblique Decision Trees for Model Explainability

Description

A surrogate model is an interpretable model trained to approximate the predictions of a black box model for model explainability[1]. Thus it should be able to approximate the predictions of the underlying model as accurately as possible and be interpretable at the same time. The decision tree model is widely used as a surrogate model since it can be easily trained and interpreted. The traditional decision tree model uses an axis-parallel split point to determine whether a data point should be assigned to the left or right branch. However, it requires many splits to achieve the approximation accuracy when decision boundaries are oblique, and, therefore, increases the interpretation difficulty. By contrast, oblique decision trees have the potential to outperform traditional decision trees because with a smaller number of splits an oblique hyperplane can achieve better separation of the instances of the data that belong to different classes[2], as shown in Fig. 1. However considering that the oblique decision rules are less intuitive, it is a question whether oblique rules provided in a visual analytics context would be a good way to facilitate the explainability compared to more common decision tree options[4].

In this project, we want to use visualization to facilitate the explainability of the oblique decision tree. We will compare oblique decision trees in relation to traditional decision trees in the approximation accuracy as a surrogate model.

References

[1].   Molnar, Christoph.“Interpretable machine learning. A Guide for Making Black Box ModelsExplainable”, 2019. https://christophm.github.io/interpretable-ml-book/global.html

[2].  Chaturvedi, Setu, and Sonal Patil."Oblique Decision Tree Learning Approaches-A Critical Review."International Journal of Computer Applications 82.13 (2013).

[3].   L. Zhang, J. Varadarajan, P. N. Suganthan, N. Ahuja and P. Moulin, "Robust Visual Tracking Using Oblique Random Forests," 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 5825-5834, doi: 10.1109/CVPR.2017.617.

[4].  S. van den Elzen and J. J. van Wijk, "BaobabView: Interactive construction and analysis of decision trees," 2011 IEEE Conference on Visual Analytics Science and Technology (VAST), 2011, pp. 151-160, doi: 10.1109/VAST.2011.6102453.

Details

Student

CL

Chicheng Liu

Supervisor

Anna Vilanova

Secondary supervisor

LC

Linhao Meng, Dennis Collaris

Link

Thesis