About¶

ShapFire is an automated and wrapper-based approach to feature importance ranking and feature selection based on SHAP Feature Importance Rank Ensembling (SHAPFire, stylized ShapFire).

ShapFire is built on top of Microsofts gradient boosting decision tree framework LightGBM and the SHAP (SHapley Additive exPlanations) Python library for Machine Learning (ML) model inspection and interpretation.

The ShapFire approach is motivated by the fact that highly associated features in an input dataset can affect ML model interpretability, making it hard to obtain accurate feature importance rankings.

ShapFire aims to work specifically in a setting where the input dataset contains several highly associated features that need to be assigned a globally consistent ranking that, e.g., a domain expert can further assess.

Getting Started¶

Install the development version from the git source:

pip install git+https://github.com/nicklasxyz/shapfire.git

Then see: Binary Classification Demo