Experiments with LIME

This repository contains code for experiments on how the choice of hyperparameters affects explanations of LIME.

Motivation

Despite being a popular framework for ad-hoc machine learning interpretation, LIME has several major disadvantages:

1. Explanations of similar examples might be totally different;

2. LIME explanations fidelity is low;

3. Explanation depends on the choice of LIME hyperparameters.

You may find details on these problems in my blog post on LIME framework problems: What’s Wrong with LIME

or in the following articles, which I used for writing the blog post:

1. On the Robustness of Interpretability Methods

2. A study of data and label shift in the LIME framework

Reproducibility of Results

Run

python -m virtualenv lime_experiments
source lime_experiments/bin/activate
pip install -r requirements.txt

Then run lime_experiments.ipynb notebook.