Implementation of ICLR 2023 Workshop paper: CGXplain: Rule-Based Deep Neural Network Explanations Using Dual Linear Programs
Accepted at ICLR Workshop on Trustworthy ML for Healthcare (TML4H).
Run the CLI commands to install required dependencies. Note that this setup requires conda, but we recommend
using mamba for a faster installation.
conda env create -f environment.yml
conda activate cgxTrain a basic neural network. We currently support tensorflow, but plan to add PyTorch support soon.
We provide some training utils, but any tensorflow model can be used to extract rules from.
import numpy as np
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from cgx.models.fcnn import train_dnn, model_fn
from cgx.explain import column_generation_rules
X, y = load_breast_cancer(return_X_y=True, as_frame=True)
x_train, x_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
model, metrics = train_dnn(
x_train=x_train,
y_train=y_train,
x_test=x_test,
y_test=y_test,
)
print(model.summary())
train_predictions = np.argmax(model.predict(x_train), axis=1)
rules = column_generation_rules(
x_train,
train_predictions,
cnf = False,
lambda0 = 0.001, # penalty for # rules
lambda1 = 0.001, # penalty for # terms
)Output
'rules': ['worst radius <= 14.97',
'concave points error > 0.01 AND worst concavity <= 0.15',
'mean compactness > 0.06 AND worst texture <= 27.83 AND worst area <= 862.07 AND worst smoothness <= 0.15 AND worst symmetry <= 0.33',
'mean radius > 13.30 AND mean fractal dimension > 0.06 AND area error <= 24.72 AND worst smoothness > 0.11 AND worst symmetry <= 0.30']