📄 中文摘要
解释性互动学习（XIL）使用户能够通过对模型解释的反馈来引导机器学习（ML）中的模型训练，从而帮助模型关注用户视角下与预测相关的特征。本研究探索了这一学习范式在缓解视觉分类器中的偏见和虚假相关性方面的能力，特别是在性别分类等易受数据偏见影响的场景中。研究考察了两种方法论上不同的最先进的XIL策略，即CAIPI和“正确理由的正确性”（RRR），以及一种结合这两种策略的新型混合方法。通过比较分割掩膜与解释的结果，定量评估了这些方法的效果。
用于视觉性别分类的解释性互动机器学习偏见缓解
📄 English Summary
Explanatory Interactive Machine Learning for Bias Mitigation in Visual Gender Classification
Explanatory interactive learning (XIL) allows users to guide model training in machine learning (ML) by providing feedback on model explanations, thus helping the model focus on features relevant to predictions from the user's perspective. This study explores the capability of this learning paradigm to mitigate bias and spurious correlations in visual classifiers, particularly in data-biased scenarios such as gender classification. Two methodologically distinct state-of-the-art XIL strategies, namely CAIPI and Right for the Right Reasons (RRR), are investigated, along with a novel hybrid approach that combines both strategies. The results are quantitatively evaluated by comparing segmentation masks with explanations.