Explanatory Interactive Machine Learning for Bias Mitigation in Visual Gender Classification

1. Problem Statement

Machine Learning (ML) models, particularly in visual classification tasks like gender recognition, frequently suffer from data bias and spurious correlations. Instead of learning relevant features (e.g., facial structure), models often rely on irrelevant contextual cues (e.g., background objects, clothing, or stereotypical associations) to make predictions. This leads to unfair outcomes and poor generalization.

While traditional bias mitigation techniques exist, they often operate in a "black-box" manner. Explanatory Interactive Learning (XIL) offers a promising alternative by integrating human feedback directly into the training loop via model explanations. However, it remains unclear how effectively different XIL strategies can detect and mitigate specific biases (like gender stereotypes) and force models to focus on semantically relevant image regions.

2. Methodology

The study proposes a systematic investigation of XIL strategies to guide models toward relevant features and away from biased ones.

A. System Architecture & Workflow

The workflow involves:

1. Baseline Training: Training a classification model on a dataset.
2. Sample Selection: Selecting specific samples for interaction using two strategies:
   • Uncertainty Sampling: Samples where the model is unsure (low confidence).
   • High-Confidence Sampling: Samples where the model is correct and confident (>90%).
3. Explainability: Generating attribution maps using GradCAM (post-hoc) or Bounded Logit Attention (BLA) (self-explaining, integrated into the model).
4. Feedback Integration: Using ground-truth segmentation masks (simulating user feedback) to steer the model via three distinct methods.

B. Feedback Integration Strategies

The paper evaluates three approaches:

1. CAIPI (Counterfactual Augmentation for Interactive Learning):
   • Generates $k$ synthetic counterexamples per sample.
   • Applies image transformations (inversion, posterization, color jittering, etc.) only to regions marked as irrelevant by the user (background), while preserving the relevant region (person).
   • This forces the model to learn that changes in the background should not alter the prediction.
2. RRR (Right for the Right Reasons) Loss:
   • Adds an auxiliary term to the loss function.
   • Penalizes the model if it relies on gradients in regions marked as irrelevant (background).
   • Formula: $L = L_{\text{cross-entropy}} + \lambda_1 L_{\text{right-reasons}} + \lambda_2 L_{\text{regularization}}$.
3. Hybrid Approach:
   • Combines both strategies: The dataset is augmented using CAIPI, and the model is trained using the RRR loss function. This leverages both data augmentation and explicit loss penalization.

C. Experimental Setup

• Dataset: A manually curated subset of MS COCO containing 1,830 images (915 male, 915 female) with pixel-accurate "person" segmentation masks.
• Models: Six pre-trained CNNs (DenseNet121, EfficientNet-B0, GoogLeNet, MobileNet-V2, ResNet50, VGG16). EfficientNet-B0 was selected as the primary baseline due to superior performance.
• Metrics:
  • Classification Accuracy.
  • Bias Metrics: Foreground Focus Proportion (FFP), Background Focus Proportion (BFP), Background Saliency Ratio (BSR), and DICE Score (alignment between model attention and ground truth).
  • Fairness: Comparison of misclassification rates between male and female classes.

3. Key Contributions

1. Systematic Investigation of XIL for Bias: The study rigorously tests how XIL strategies (CAIPI, RRR, and Hybrid) mitigate gender bias in image classifiers.
2. Novel Hybrid Strategy: Proposes a combined approach integrating CAIPI (data augmentation) and RRR (loss function) to optimize models for both accuracy and fairness.
3. Qualitative & Quantitative Analysis: Evaluates model learning using both post-hoc (GradCAM) and self-learned (BLA) explainability methods to verify if models are focusing on relevant regions.
4. Insight on Sampling Strategies: Challenges the conventional wisdom of Active Learning by demonstrating that High-Confidence Sampling is often more effective than Uncertainty Sampling for XIL bias mitigation.

4. Key Results

• Focus on Relevant Features:
  • CAIPI and the Hybrid approach significantly improved the model's focus on the foreground (the person).
  • The Hybrid approach achieved the best reduction in reliance on background features (lowest BFP and BSR).
  • RRR showed improvements but was less pronounced than CAIPI-based methods.
• Bias Mitigation & Fairness:
  • All XIL methods reduced the disparity in misclassification rates between males and females.
  • The baseline model misclassified females 53.8% of the time vs. 46.2% for males. XIL methods balanced this distribution (e.g., CAIPI High-Confidence $k=5$ achieved 52.0% vs. 48.0%).
• Accuracy vs. Fairness Trade-off:
  • Generally, bias mitigation led to a slight decrease in overall classification accuracy.
  • Exception: CAIPI with Uncertainty Sampling ($k=1$) actually improved accuracy (from 74.58% to 75.42%), proving that fairness and accuracy can be mutually beneficial.
• Sampling & Explainability:
  • High-Confidence Sampling consistently outperformed Uncertainty Sampling in XIL contexts.
  • BLA (self-explaining) generally yielded better results and alignment than GradCAM (post-hoc).

5. Significance and Conclusion

This paper demonstrates that Explanatory Interactive Learning is a viable and effective strategy for mitigating data bias in visual classifiers. By allowing models to "see" what humans consider irrelevant (via segmentation masks) and adjusting training accordingly, models can be steered away from spurious correlations like gender stereotypes.

• Practical Impact: The findings suggest that integrating human feedback on why a model made a decision is more powerful than just correcting the label.
• Methodological Insight: The study highlights that CAIPI is particularly effective at forcing models to ignore background noise, while the Hybrid approach offers the most robust bias reduction.
• Future Work: The authors note that while simulations using ground-truth masks were successful, future research must validate these methods with real human users and explore their application to multi-label and multi-class problems.

In summary, XIL provides a pathway to more transparent, fair, and robust AI systems by aligning model reasoning with human expectations, even if it occasionally incurs a minor cost in raw classification accuracy.