Test-fairness deep learning with influence score

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• AI performance disparities can arise from training data collected from specific populations, leading to unequal outcomes across sensitive groups or data sources.
• Ensuring fairness and equitable performance in AI models across diverse datasets is a critical challenge in machine learning.

Purpose of the Study:

• To propose and evaluate a feature-selection-based method for improving test-fairness in deep learning models.
• To reduce cross-dataset subgroup performance disparity while preserving overall prediction performance.

Main Methods:

• Developed a feature-selection approach leveraging the influence score (I-score) to identify and mask dataset-associated features.
• Trained an auxiliary model to predict dataset origin and used I-score to select features strongly linked to dataset membership.
• Applied the method to deep learning models for skin lesion classification using two distinct datasets (ISIC 2019 and ASAN).

Main Results:

• The proposed fair I-score model successfully reduced cross-dataset subgroup performance disparity.
• High classification performance for skin lesion prediction was maintained after applying the fairness intervention.
• Empirical results demonstrated the effectiveness of the method on real-world skin lesion datasets from different populations.

Conclusions:

• Feature selection using the influence score is an effective strategy for enhancing test-fairness in AI models.
• The proposed method offers a practical solution for mitigating performance disparities in AI systems trained on heterogeneous data.
• This approach contributes to the development of more equitable and reliable AI applications, particularly in medical image analysis.