README.md

metadata

title: SHAP Explainability Demo
emoji: 🔍
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: mit

SHAP Explainability Demo 🔍

An interactive demonstration of SHAP (SHapley Additive exPlanations) algorithm with three different explanation approaches.

🎯 Features

This demo showcases three powerful SHAP explanation methods:

1. 🖼️ Pixel-level Explanations (MNIST Digits)

• Uses DeepExplainer for deep learning models
• Explains which pixels contribute to digit classification
• Interactive digit selection (0-9)
• Real-time visualization with red/blue attribution maps

2. 🎨 Image Segmentation Explanations (ImageNet)

• Uses Partition Explainer with image masking
• Explains image classification with region-based attributions
• Upload any image and see which regions matter
• Shows top 4 predicted classes with real names (e.g., "beagle", "golden_retriever")
• Uses ResNet50 pre-trained on ImageNet

3. 📊 Tabular Data Explanations (Adult Income)

• Uses TreeExplainer for tree-based models
• Explains income prediction with feature attributions
• Waterfall plots showing feature contributions
• Based on Adult Income dataset with Random Forest classifier

🚀 How to Use

Tab 1: Pixel-level (MNIST)

1. Use the slider to select a digit index (0-99)
2. Click "Generate Explanation"
3. See the original digit and SHAP pixel attributions

Tab 2: Image Segmentation (ImageNet)

1. Upload any image (JPG, PNG)
2. Click "Generate Explanation"
3. Wait 30-60 seconds (image masking is computationally intensive)
4. See SHAP region attributions for top 4 classes

Tab 3: Tabular Data (Adult Income)

1. Use the slider to select a sample (0-99)
2. Click "Generate Explanation"
3. See the waterfall plot showing feature contributions

🛠️ Technologies

• Gradio: Web interface
• SHAP: Explanation framework
• PyTorch: MNIST model
• TensorFlow/Keras: ResNet50 model
• scikit-learn: Random Forest model
• OpenCV: Image inpainting for masking

📚 About SHAP

SHAP (SHapley Additive exPlanations) is a game-theoretic approach to explain machine learning model predictions.

Learn more:

• SHAP GitHub
• SHAP Paper

📄 License

MIT License

---

Built with ❤️ using Gradio and SHAP