Toward Understanding the Disagreement Problem

This repository provides the code to reproduce results in the paper "Toward Understanding the Disagreement Problem in Neural Network Feature Attribution" submitted to the 2nd World Conference on eXplainable Artificial Intelligence (XAI).

The structure is as follows:

• Sec_1_Introduction.R reproduces Figure 1 in the introduction, which shows the disagreement of popular feature attribution methods on the COMPAS dataset.

• Sec_3_Understanding_the_Distribution.R reproduces the figures from the running example in Section 3.

• Sec_4_Simulations.R runs the simulation study (can last up to several days) and creates the figures in Section 4.

• Sec_Appendix.R creates the Figure 7 showing the transformations for the non-linear effects.

All created figures are saved in the figures/ folder. However, the final figures used in the paper and the results from the simulation study of Section 4 are stored in the results/ folder.