Welfare Sensitive Machine Learning: Applications and Experiments with Consumer Lending in Kenya

An increasing number of decisions are guided by machine learning algorithms. This is old news in the Bay Area, but is only a recent trend in international development, where algorithms are now being used to direct humanitarian aid and facilitate loans to the ultra-poor. This talk will provide an overview and preliminary results from two studies that develop new approaches to machine learning that are designed for deployment in low-resource settings.

The first study explores the potential for “welfare-sensitive” machine learning, in which the goal of the machine learning algorithm is not just to maximize predictive performance, but also to maximize the social welfare impact of the resulting decision rule. I will sketch a formal theory of multi-objective optimization that treats social welfare as a primitive objective, which can be predicted from data and algorithmically optimized.

The second study proposes a framework for “manipulation-proof” machine learning that is designed to address two common issues that arise in machine learning deployments. First, when algorithms are being used to allocate scarce resources, this creates incentives for people to act strategically to game the algorithm. Second, consumers want to know how these decisions are being made — but disclosing the decision rule makes it easier to game the system. I will describe a new class of estimators that are stable under manipulation, even when the decision rule is fully transparent. I will also present pilot results from a large field-based study in Nairobi that shows that this strategy-robust estimator outperforms standard approaches to supervised learning.