Ginger Anchors

Implementation and Extensions of Anchors: High-Precision Model-Agnostic Explanations. The re-implementation was done in the iML lecture (WS21/22).

Contributions

• Implementation of Anchors via the Bottom-up construction approach.
• Implementation of Beam Search on top of the Bottom-up construction.
• Simple interfaces with well-splitted main functions.
• Analysis on how 𝐵, 𝛿 and 𝜖 influence the results.
• SMAC3 as alternative anchor finder.

References

• Paper Reference: https://homes.cs.washington.edu/~marcotcr/aaai18.pdf
• Code Reference: https://github.com/marcotcr/anchor

Installation

Note: swig is needed to install SMAC3. See installation instructions.

Create a conda environment

$ conda create -n GingerAnchors python=3.9
$ conda activate GingerAnchors
$ pip install ginger-anchors

Usage

You can get an explanation by setting up an Explainer and calling one of three search functions.

exp = Explainer(X_df)
anchor = exp.explain_bottom_up(instance, model, tau=0.95)
print(anchor.get_explanation())

For a more detailed example, see src/main.py.

Analysis

The plots were too large to put them into this repository. Please download them from seafile. To reproduce the raw data, run:

ginger-anchors> python src/analysis.py

A preview can be found in our writeup: analysis.md

Authors

Jim Rhotert & Julian Bilsky