ExplainerPFN

ExplainerPFN is a novel explainability method. This project provides fast, zero-shot feature importance estimation without requiring access to the original prediction model, making it ideal for model-agnostic explainability in tabular data scenarios.

Key Features

• Zero-shot Explainability: Predict feature contributions without retraining or access to the original model
• Fast Inference: Leverages TabPFN's efficient transformer architecture for rapid explanations
• SHAP-compatible: Uses SHAP values as ground truth for validation and benchmarking
• Model-agnostic: Works with any black-box model's predictions
• Customizable Corrections: Multiple correction methods (statistical, linear, multiplicative, additive) to improve explanation quality
• Fine-tuning Support: Ability to fine-tune the foundation model on custom datasets
• Comprehensive Metrics: Built-in evaluation metrics for explanation quality (fidelity, consistency, sensitivity)

Installation

# Clone the repository
git clone https://github.com/yourusername/ExplainerPFN.git
cd ExplainerPFN

# Install dependencies
pip install -r requirements.txt

Requirements

• Python 3.8+
• NumPy
• pandas
• scikit-learn
• PyTorch
• TabPFN 2.1.2
• SHAP
• matplotlib
• tqdm

Quick Start

Basic Usage

from explainerpfn.model import ExplainerPFN
import numpy as np

# Initialize the explainer
explainer = ExplainerPFN(
    n_estimators=1,
    device="auto",  # Automatically selects GPU/CPU
    fit_mode="fit_with_cache"
)

# Fit on background data
# X: feature matrix (n_samples, n_features)
# y: target values or model predictions
explainer.fit(X_train, y_train)

# Get feature importance explanations
explanations = explainer.predict(X_test, y_test)

# Apply correction (optional but recommended)
corrected_explanations = explainer.apply_correction(
    y_test, 
    explanations, 
    kind=["statistical", "additive"]
)

Fine-tuning on Custom Data

Note: Fine-tuning was working at some point but may require adjustments since some things were modified in the meantime and we haven't tested this method since.

import torch.optim as optim

# Prepare optimizer
optimizer = optim.Adam(explainer.model_.parameters(), lr=1e-4)

# Fine-tune with SHAP values as ground truth
for epoch in range(num_epochs):
    loss = explainer.finetune(optimizer, X_train, y_train, shap_values)
    print(f"Epoch {epoch}: Loss = {loss}")

# Save the fine-tuned model
explainer.save_foundation_model("path/to/model.ckpt")

Architecture

ExplainerPFN follows the sequence below for generating feature importance explanations:

Input (X, y, feature_idx)
    ↓
Preprocessing Pipeline
    ├── Categorical encoding
    ├── Feature distribution reshaping
    └── Target transformations
    ↓
TabPFN Transformer Encoder
    ↓
Bar Distribution Logits
    ↓
Statistical Corrections
    ↓
Feature Importance Values

Key Components

• ExplainerPFN: Main class implementing the explainer interface
• Inference Engine: Efficient batched inference with caching
• Preprocessing: Customizable feature transformations and ensembling
• Metrics Module: Evaluation tools for explanation quality
• Training Module: Synthetic data generation for pre-training

Research & Experiments

The notebooks/ directory contains Jupyter notebooks demonstrating:

1. SHAP value estimation: Comparison with traditional SHAP methods
2. Cross-dataset evaluation: Performance across multiple tabular datasets
3. Ablation studies: Component-wise analysis of the method
4. Case studies: Real-world application examples
5. Synthetic data generation: DAG-based synthetic data for training
6. Model multiplicity analysis: Understanding explanation variance

Project Structure

ExplainerPFN/
├── explainerpfn/           # Main package
│   ├── base.py            # Core ExplainerPFN class
│   ├── inference.py       # Inference engine implementation
│   ├── preprocessing.py   # Data preprocessing utilities
│   ├── config.py          # Configuration settings
│   ├── metrics/           # Evaluation metrics
│   ├── model/             # Model architecture components
│   ├── train/             # Training utilities and data generation
│   └── visualization/     # Plotting and visualization tools
├── notebooks/             # Example notebooks and experiments
├── requirements.txt       # Package dependencies
├── Makefile              # Utility commands
└── README.md             # This file

Citation

If you use ExplainerPFN in your research, please cite:

@inproceedings{
    TBD
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Built on top of TabPFN by Prior Labs
• Uses SHAP for ground truth explanations