Explanatory Interactive Machine Learning for Bias Mitigation in Visual Gender Classification

This repository contains the implementation, dataset, and experiments for the paper:
“Explanatory Interactive Machine Learning for Bias Mitigation in Visual Gender Classification.”

We explore two state-of-the-art Explanatory Interactive Learning (XIL) strategies — CAIPI and Right for the Right Reasons (RRR) — and propose a novel hybrid approach to guide visual classifiers toward more relevant features, reducing bias and improving fairness in gender classification tasks.

---

🚀 Features

• Implementation of CAIPI and RRR XIL strategies
• Novel Hybrid XIL approach combining strengths of both methods
• Evaluation of fairness, transparency, and interpretability in visual gender classification
• Comparisons using GradCAM and Bounded Logit Attention (BLA) for model explanations
• Experiments showing how XIL can reduce bias (e.g., balancing misclassification rates between male and female predictions)

---

📦 Installation

Clone the repository:

git clone https://github.com/yourusername/xil-bias-gender-classification.git
cd xil-bias-gender-classification

pip install -r requirements.txt

---

📊 Experiments

We evaluate models on:

• Baseline classifiers without XIL
• CAIPI with varying numbers of augmentations
• RRR with explanation-based regularization
• Hybrid combining CAIPI and RRR
• 2 explainability methods GradCAM and Bounded Logit Attention (BLA)
• 2 sampling strategies Uncertainty sampling High Confidence sampling

---

📂 Dataset

The dataset used in this study is a subset of the MS COCO dataset that contains images of persons.
Since MS COCO does not natively provide annotations for gender or sex, a manual data selection process was carried out to construct a binary gender classification dataset.

⚠️ Note: We are aware that binary gender represents an over-simplification. To limit task complexity and simplify result interpretation, the scope of this study was restricted to binary classification.

---

📖 Citation

If you use this work, please cite:

  @article{tbd
  }