Inclusive Decision Making via Contrastive Learning and Domain Adaptation

Abstract

Inclusive decision-making is crucial for promoting social justice and welfare, particularly in high-stakes scenarios like loan screening. We propose using contrastive learning and domain adaptation to improve inclusion in algorithmic decision-making. Specifically, we focus on micro-lending, which has played a significant role in increasing access to financial services. Traditional machine learning algorithms for credit evaluation face the selective labels problem, as the training data usually only contains default outcome labels from approved loan applications that tend to represent borrowers with more favorable socioeconomic characteristics. Consequently, these algorithms struggle to effectively generalize to disadvantaged borrowers. To overcome this problem, we introduce a Transformer-based loan screening model that leverages self-supervised contrastive learning and domain adaptation. Our model uses contrastive learning to train our feature extractor on unapproved (unlabeled) loan applications and employs domain adaptation to generalize the performance of our label predictor. We evaluate our approach on a real-world micro-lending dataset and demonstrate its effectiveness. The results show that our approach significantly enhances inclusion in funding decisions, while simultaneously improving loan screening accuracy and lender profit by 7.10% and 8.95%, respectively. Additionally, we find that incorporating test data and labeling a small ratio of it further enhances model performance.

Keywords: Contrastive Learning, Domain Adaptation, FinTech, Inclusion, Representation Bias