Class Contrastive Explanations and Generating Complex Stories from Machine Learning Models

In order to provide a better explanability for machine learning models in healthcare are, we provide an implementation for class contrastive analysis for Pima dataset.

Class contrastive analysis is a technique to explain black-box machine learning models [1].

In addition to this approach, Mark W. Craven and Jude W. Shavlik proposed extracting Thee-Structured representations of trained networks in 1993.

We use the TREPAN implementation in "Extracting tree-structured representations of trained networks" : Craven,Shavlik 1993 from github user: @abarthakur.

https://github.com/abarthakur/trepan_python

Quick Installation

Install Anaconda (Python distribution)

and install all packages by running the following command at the Terminal

pip install -r requirements.txt

Data

Download the data from

https://www.kaggle.com/datasets/uciml/pima-indians-diabetes-database

Files

The script simple_2D__pima_heatmap.ipynb does the following things:

1. Trains a simple 3-layer neural network on the pima dataset
2. Provides Data centric explanation like outlier detection
3. Provides double feature class contrastive analysis

Figure 3.1 is generated in pima_heatmap.ipynb.

Figure 4.1 is generated from pima_heatmap.ipynb.

Figure 4.2 is generated from 2D_pima_heatmap.ipynb.

Figure 4.3 is generated from 3D_pima_heatmap.ipynb.

Figure 4.4 is generated from 4D_pima_heatmap.ipynb.

Figure 4.7 is generated from 2D_pima_heatmap.ipynb.

Figure 4.8 is generated from opposite2D_pima_heatmap.ipynb.

Manuscript and citation

Banerjee S, Lio P, Jones PB, Cardinal RN (2021) A class-contrastive human-interpretable machine learning approach to predict mortality in severe mental illness. npj Schizophr 7: 1–13.

https://www.nature.com/articles/s41537-021-00191-y

Generating complex explanations from machine learning models using class-contrastive reasoning

https://www.medrxiv.org/content/10.1101/2023.10.06.23296591v1

Contact

Yujia Yang and Soumya Banerjee

[email protected]

References

If you use this work, please cite the following paper:

1. Banerjee S, Lio P, Jones PB, Cardinal RN (2021) A class-contrastive human-interpretable machine learning approach to predict mortality in severe mental illness. npj Schizophr 7: 1–13.

https://www.nature.com/articles/s41537-021-00191-y

2. Yujia Yang, Soumya Banerjee, Generating complex explanations from machine learning models using class-contrastive reasoning

https://www.medrxiv.org/content/10.1101/2023.10.06.23296591v1