Collusion in Knowledge Elicitation for Lending: Deep Learning for Colluders and Collusion Detection

Abstract

Microfinance is an increasingly important tool for poverty alleviation, but it is difficult for lenders to discern the creditworthiness of potential borrowers with limited credit history. To address this challenge, York et al. constructed elicitation mechanisms that incentivize community members to give truthful reports on potential borrowers when reporting independently. We extend the York et al. model by allowing recommenders to collude with one another and have intrinsic preferences over borrowers. Our model of recommender utility combines profit in the scoring rule literature and intrinsic preferences in the social choice literature, thus connecting the problems of knowledge elicitation for lending and social choice. We use deep learning to find collusive strategies for recommenders, which we believe to be a novel approach to studying collusion in the context of decision scoring rules. We develop methods of statistical machine learning for detecting collusion, and we test their efficacy on the learned collusive strategies. Lastly, we study the coevolutionary dynamic between colluders and collusion detectors using generative adversarial networks (GANs).