SparseExplanationValues

The introduction of SEV has provided an sparse explanation for each unit based on a binary machine learning classifier and new ways to increase this sparsity explanation without its performance (like Accuracy and AUC). This code is used for AISTATS 2024, which contains implementations and Experiments for Sparse Explanations Values (SEV) and two optimal methods: All-Opt and Vol-Opt.

Development

Guidelines for developers who wants to reproduce the experiments results, modify and test the code.

Repository Structure

• SEV: The implementation of the definition of Sparse Explanations and two optimization methods.

It consists 5 files. The SEV.py file is the implementation of different SEV definitions: SEV+, SEV- and SEV$^{\circledR}$. Encoder.py file the data encoder for doing one-hot encoding for the general datasets, save the median values for numerical features and mode values for categorical features. OptimizedSEV.py includes all the optimization methods: AllOpt+ for optimizing SEV+, AllOpt- for optimizing SEV-, AllOptR for optimizing SEV$^{\circledR}$, and VolOpt for optimizing SEV+ in linear classifiers. It also includes the models used in the experiments: Logisitic Regression(SimpleLR), 2-layer MLP(SimpleMLP) and Graident Boosting Decision Trees(GBDT).

• Plots: The code for generating figures in the paper

• Experiments: The code for different experiments

The expeirments folder consists of two parts: Code and Scripts. The code folder contains thepython files for different experiments and scripts folder shows how to submit the experiments through slurm.

General Usage of SEV

The usage of SEV can reference the Experiment Part of SEV. The general usage of SEV can be used as follows:

from SEV.SEV import SEVPlot, SEV
from SEV.data_loader import data_loader
from SEV.Encoder import DataEncoder
from sklearn.linear_model import LogisticRegression

# get the dataset and the negative population
X,y,X_neg = data_loader(args.dataset)

# preprocessing the dataset
encoder = DataEncoder(standard=True)
# fit the encoder with the negative population
encoder.fit(X_neg)
# transform the whole dataset
encoded_X = encoder.transform(X)

# construction the model
lr = LogisiticRegression()
lr.fit(encoded_X,y)

# for explaining the whole dataset, "plus" for SEV+, "minus" for SEV-
SEVPlot(lr,encoder, encoded_X, "plus")

# for explaning the one single instance
sev = SEV(lr,encoder,encoded_X.columns)
# get the number of sev for this instance
sev_num = sev.sev_cal(np.array(encoded_X.iloc[0]).reshape(1,-1),mode="plus")
print("The SEV Value for instance 0 is %d."%sev_num)
# get the features can be used in this explanation
features = sev.sev_count(np.array(encoded_X.iloc[0]).reshape(1,-1),mode="plus",choice=sev_num)
print("The feature used in this explanation are %s."%features)

The use of different optimization methods can be seen in Experiment 1 in Experiment Folder.

Requirements

The following dependencies need to be installed to run the experiments.

numpy==1.23
pandas==1.5.2
scikit-learn==1.2.2
pytorch==2.0.0
matplotlib==3.7.1
seaborn==0.12.2
psankey==1.0.1
dice-ml==0.9
shap==0.40.0
lime==0.2.0

Note: The sankeyplots are generated based on the psankey packages (https://github.com/mandalsubhajit/psankey).

Citation

To use our method please cite the paper:

@article{sun2024sparse,
  title={Sparse and Faithful Explanations Without Sparse Models},
  author={Sun, Yiyang and Chen, Zhi and Orlandi, Vittorio and Wang, Tong and Rudin, Cynthia},
  journal={Society for Artificial Intelligence and Statistics (AISTATS)},
  year={2024}
}