CLE-SH: Comprehensive Literal Explanation Package for SHapley Values by Statistical Validity

CLE-SH is a Python library designed to simplify the interpretation of SHAP values through statistical validation. By integrating feature selection, univariate analysis, and interaction analysis into a unified automated pipeline, it bridges the gap between model explainability and statistical rigor.

Updates

• JAN 28, 2026: 1.0.0 release
• JAN 16, 2026: CLE-SH is published on IEEE ACCESS

Installation

pip install cle-sh

Requirements

• Python >= 3.8
• numpy >= 1.20.0
• pandas >= 1.3.0
• matplotlib >= 3.3.0
• seaborn >= 0.11.0
• scipy >= 1.7.0
• scikit-learn >= 1.0.0
• shap >= 0.40.0
• fpdf2 >= 2.7.0

Quick Start

import pandas as pd
import numpy as np
from clesh import Explainer, CLEConfig

# Load your data
X = pd.read_csv("features.csv")
shap_values = np.load("shap_values.npy")

# Initialize with default configuration
explainer = Explainer(X=X, shap_values=shap_values)

# Perform complete analysis
results = explainer.analyze()

# Save plots
explainer.save_plots("./output")

# Generate PDF report
explainer.generate_report("./report.pdf", label="My Dataset")

# Manual analysis on specific features
univariate_result = explainer.univariate("feature_name")
interaction_result = explainer.inter("target_feature", "interaction_feature")

Configuration

All analysis parameters are configured using the CLEConfig dataclass:

from clesh import CLEConfig

config = CLEConfig(
    cont_bound=10,              # Threshold for determining continuous features
    candidate_num_min=10,        # Minimum number of features to select
    candidate_num_max=20,        # Maximum number of features to select
    p_feature_selection=0.05,    # P-value threshold for feature selection
    manual_num=0,               # Manual override (0 = automatic)
    p_univariate=0.05,          # P-value threshold for univariate analysis
    p_interaction=0.05          # P-value threshold for interaction analysis
)

All parameters have sensible defaults.

API Reference

Main Classes

Analyzer

Core analysis engine that performs all statistical calculations.

from clesh import Analyzer

analyzer = Analyzer(X, shap_values, config)
results = analyzer.analyze()

# Manual analysis
univariate = analyzer.univariate("feature_name")
interaction = analyzer.inter("target_feat", "interaction_feat")

Explainer

High-level coordinator

explainer = Explainer(X=X, shap_values=shap_values, config=config)
results = explainer.analyze()
explainer.save_plots("./output")
explainer.generate_report("./report.pdf", label="Dataset")

AnalysisResults

Dataclass containing all analysis results:

• feature_importance: DataFrame with feature rankings
• selected_features: List of selected feature names
• feature_types: Dict mapping feature names to types
• univariate_results: Dict of univariate analysis results
• interaction_results: Dict of interaction analysis results
• best_functions: Dict of best fitting functions for continuous features

Visualization Module

All plotting functions are in clesh.visuals and return figure/axis objects:

from clesh.visuals import (
    plot_shap_summary,
    plot_discrete_univariate,
    plot_continuous_univariate,
    plot_discrete_interaction,
    plot_continuous_interaction,
)

fig, ax = plot_shap_summary(shap_values, X)
fig.savefig("summary.png")

Advanced Usage

Custom Feature Analysis

from clesh import Analyzer, CLEConfig

analyzer = Analyzer(X, shap_values)

# Analyze specific feature
result = analyzer.univariate("age")
print(f"Feature type: {result['feature_type']}")
print(f"Statistics: {result['statistics']}")

# Interaction analysis
inter_result = analyzer.inter("target_feature", "interaction_feature")
print(f"Significant interaction: {inter_result['statistics']['significant']}")

Programmatic Report Generation

from clesh import Explainer
from clesh.report import generate_pdf_report

explainer = Explainer(X=X, shap_values=shap_values)
results = explainer.analyze()

# Generate PDF directly
generate_pdf_report(results, "./report.pdf", "My Dataset", config)

Output Structure

Results are stored in-memory as AnalysisResults dataclass. Plots and reports are saved to specified directories:

output/
├── shap_summary_plot.jpg
├── univariate_analysis/
│   └── *.jpg
├── interaction_analysis/
│   └── *.jpg
└── clesh_report.pdf

Citation

If you use CLE-SH or ideas from the package in your research, please cite our paper:

@ARTICLE{kim2026cle-sh,
author={Kyungjin Kim and Youngro Lee and Jongmo Seo},
journal={IEEE Access},
title={CLE-SH: Comprehensive Literal Explanation Package for SHapley Values by Statistical Validity},
year={2026},
volume={14},
number={},
pages={12514-12525},
doi={10.1109/ACCESS.2026.3654890}}

Future Plans

The project is currently undergoing a refactoring process to enhance the literal explanation engine and address minor bugs.