sbobek / inxai Goto Github PK

Iterating with enumerate does not work unless y and X has indices that start from 0.

for index, (skip, row) in enumerate(X.iterrows()):
correct_label = y[index]

In order to walk-around, call reset_index(drop=True) before passing X, and y to the function.

Perform scaling before calculating inverse shares of importance for particula features/segments.

• Create a set of synthetic datasets using make_classification, make_blobs or make_circles , make_moons
• Using these datasets build classifiers using different algorithms
• Explain classifiers using LIME/SHAP/Other..
• Evaluate explanations with INXAI (try corrupt explanations and see if INXAI detects intentionally made bad explanations)
• Summarize results

Correctenss metric specific to Infomraiton retrieval and recommender systems that results are based on rankings: https://dl.acm.org/doi/10.1145/3331184.3331377

In order to make results of different explainers comparable, one need to scale it.

Note that classic per-column or per-sample scaling will corrupt the internal dependencies between importance and features.
Therefore the scaling is performed over the whole matrix of explanation generated by the same model.

Here is an updated requirements.txt file. To create conda environment from scratch with all required packages, run commands below:

conda create --name inxai -y -c conda-forge --file requirements.txt
conda activate inxai