ARTEMIS: A Robust Toolkit of Explanation Methods for Interaction Spotting

A Python package with explanation methods for extraction of feature interactions from predictive models

Overview

artemis is a Python package for data scientists and machine learning practitioners which exposes standardized API for extracting feature interactions from predictive models using a number of different methods described in scientific literature.

The package provides both model-agnostic (no assumption about model structure), and model-specific (e.g., tree-based models) feature interaction methods, as well as other methods that can facilitate and support the analysis and exploration of the predictive model in the context of feature interactions.

The available methods are suited to tabular data and classification and regression problems. The main functionality is that users are able to scrutinize a wide range of models by examining feature interactions in them by finding the strongest ones (in terms of numerical values of implemented methods) and creating tailored visualizations.

Documentation

Full documentation is available at https://pyartemis.github.io/.

Installation

Latest released version of the artemis package is available on Python Package Index (PyPI):

pip install -U pyartemis

The source code and development version is currently hosted on GitHub.

Authors

The package was created as a software project associated with the BSc thesis Methods for extraction of interactions from predictive models in the field of Data Science (pl. Inżynieria i analiza danych) at Faculty of Mathematics and Information Science (MiNI), Warsaw University of Technology in cooperation with NASK National Research Institute.

The authors of the artemis package are:

• Paweł Fijałkowski
• Mateusz Krzyziński
• Artur Żółkowski

BSc thesis and work on the artemis package was supervised by Przemysław Biecek, PhD, DSc.