Fair CCA for Fair Representation Learning: An ADNI Study

This repository contains the implementation of FR-CCA (Fair Representation Canonical Correlation Analysis), a novel method for fair representation learning that ensures projected features are independent of sensitive attributes while maintaining high correlation analysis performance.

📋 Overview

Traditional Canonical Correlation Analysis (CCA) methods often overlook fairness considerations, potentially leading to biased representations that amplify societal biases. Our FR-CCA method addresses this limitation by:

• Ensuring independence between learned representations and sensitive attributes
• Maintaining high correlation between different data modalities
• Optimizing for both fairness and classification performance
• Providing computational efficiency comparable to standard CCA

🧠 Method

FR-CCA extends traditional CCA by adding fairness constraints that prevent any classifier from predicting sensitive attributes (e.g., sex, race, age) based solely on the projected data. The method solves:

max trace(U^T X^T Y V)
subject to: U^T X^T X U = V^T Y^T Y V = I_R
           U^T X^T ẑ = V^T Y^T ẑ = 0

Where ẑ represents the centered sensitive attribute vector.

📁 Repository Structure

├── FRCCA_AV14_MRI_Update_X.ipynb    # Implementation for MRI modality
├── FRCCA_AV14_MRI_Update_Y.ipynb  # Implementation for AV1451 (Tau PET) modality
├── data/                          # Data directory (ADNI datasets)
├── utils/                         # Utility functions
└── README.md                      # This file

🔧 Requirements

pip install numpy pandas scikit-learn matplotlib seaborn
pip install scipy shap jupyter

📊 Datasets

ADNI (Alzheimer's Disease Neuroimaging Initiative)

• MRI: 66 cortical thickness features from structural MRI scans
• AV1451: 68 SUVR features from Tau PET scans
• Participants: 375 subjects (182 males, 193 females)
• Diagnoses: Cognitively Normal (CN), Mild Cognitive Impairment (MCI), Alzheimer's Disease (AD)

Synthetic Data

• Multivariate Gaussian distributions with embedded sensitive attribute bias
• 500 samples with controllable correlation structure
• Used for method validation and parameter tuning

🚀 Usage

For Synthetic X Modality:

# Run the synthetic X notebook
jupyter notebook FRCCA_Synthetic_X.ipynb

For Synthetic Y Modality:

# Run the synthetic T notebook
jupyter notebook FRCCA_Synthetic_Y.ipynb

Need access to ADNI data when running MRI and AV14 notebook

For MRI Modality:

# Run the MRI notebook
jupyter notebook FRCCA_AV14_MRI_Update_X.ipynb

For AV1451 (Tau PET) Modality:

# Run the AV1451 notebook
jupyter notebook FRCCA_AV14_MRI_Update_Y.ipynb

Both notebooks include:

1. Data preprocessing and standardization
2. FR-CCA model training
3. Fairness evaluation (DPG, EOG, GSG metrics)
4. Classification performance assessment

📈 Key Results

Our FR-CCA method demonstrates:

• Superior Fairness: Significant reduction in demographic parity gap (DPG), equalized odds gap (EOG), and group sufficiency gap (GSG)
• Maintained Accuracy: Competitive classification performance compared to traditional methods
• Computational Efficiency: Time complexity similar to standard CCA: O((Dx + Dy)³ + N·Dx² + N·Dy² + N·Dx·Dy)
• Clinical Relevance: Identifies important brain regions consistent with Alzheimer's disease pathology

Performance Comparison

Method	GSG ↓	DPG ↓	EOG ↓	Accuracy
SVM	0.329±0.042	0.081±0.040	0.084±0.037	0.612±0.033
CCA	0.239±0.208	0.028±0.024	0.015±0.017	0.635±0.026
SF-CCA	0.286±0.033	0.036±0.024	0.033±0.031	0.619±0.026
MF-CCA	0.242±0.042	0.015±0.013	0.026±0.022	0.621±0.015
FR-CCA	0.215±0.041	0.008±0.006	0.010±0.005	0.612±0.009

🔬 Key Features

Fairness Metrics

• Demographic Parity Gap (DPG): Measures score consistency across sensitive groups
• Equalized Odds Gap (EOG): Ensures equal treatment across all label-attribute combinations
• Group Sufficiency Gap (GSG): Validates prediction relevance regardless of sensitive attributes

Brain Region Analysis

• MRI Focus: Memory, language, and visual processing regions (entorhinal cortex, temporal gyri, fusiform gyrus)
• AV1451 Focus: Tau pathology in sensory processing and emotional regulation areas (postcentral gyrus, temporal pole, anterior cingulate)

📝 Citation

@misc{hou2025fairccafairrepresentation,
      title={Fair CCA for Fair Representation Learning: An ADNI Study}, 
      author={Bojian Hou and Zhanliang Wang and Zhuoping Zhou and Boning Tong and Zexuan Wang and Jingxuan Bao and Duy Duong-Tran and Qi Long and Li Shen},
      year={2025},
      eprint={2507.09382},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2507.09382}, 
}

🤝 Contributing

We welcome contributions! Please feel free to submit issues, feature requests, or pull requests.

📄 License

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

🙏 Acknowledgments

• Alzheimer's Disease Neuroimaging Initiative (ADNI) for providing the neuroimaging data
• The research community for foundational work in fair machine learning and canonical correlation analysis

📞 Contact

For questions or collaboration opportunities, please open an issue in this repository.

---

Note: This implementation is for research purposes. Clinical applications should undergo appropriate validation and regulatory approval.