ICEPD can be used to detect interactions using partial dependence and produce PD and ICE plots to visualize non-linearity.

We recommend you look at the tutorial in the Examples folder (Example_icepd.ipynb) for an overview of the important functionality.

Features:

Partial dependence (PD):
- PD plots for 1D and 2D
- Ensemble options for 1D plots to show uncertainity
- Gradient of PD for 1D plots
- Linear model comparison for 1D plots
- PD plots for 1D One Hot Encoded variables
- Distributions of variable shown

Individual Conditional Expectation Plots (ICE): 
- ICE plots for 1D
- Gradients available
- Ensemble option

Interactions:
- H-statistics using PD
- H-statistics using PD for null distribution
- Statistical significance interaction detection
- LASSO interaction detection
- Bidirectional stepwise detection (not recommended)
- GUIDE algorithm interaction detection (not recommended when collinearity present)


Installation Guide

Python version 3.7 is required.

Unix

This code has been tested with Ubuntu 20.04 and Python version 3.7.

For the development version:
git clone https://github.com/aa840/icepd.git
cd icepd
pip install .

The installation process should take less than 2 minutes. 

Citation

To cite the program please use:
Machine learning of material properties: Predictive and interpretable multilinear models, Allen and Tkatchenko, Sci. Adv. 8, eabm7185 (2022)

Author: Alice E A Allen