Fairness in Machine Learning: Investigating Gender Bias in Fintech Non-Mortgage Home Loans

Abstract

Problem definition: The United States, among other countries, prohibits the collection and use of protected features such as gender, race, and age when making credit decisions - including home credit. This is done in order to explicitly prevent discrimination based on such features. However, other countries do permit the collection of this data, and it is unclear which approach is better, if the goal is to prevent bias, especially with the introduction of machine learning algorithms. I look at a large data set, collected in countries where the collection of gender data is permitted, to see 1. if there is gender bias present when machine learning models are trained with gender included as a variable, 2. if there is gender bias present when machine learning models are trained with gender excluded as a variable, and 3. if there is a significant difference between the two results. Relevance: Financial firms that rely heavily on technology are becoming more prevalent, and the decisions their algorithms make regarding home loans are critical. Further, in general, bias in machines and algorithms are important to identify because this type of discrimination is often hidden behind layers of abstraction, despite efforts to reduce bias. Methodology: I use logistic regression, random forest, and gradient boosting models to classify applicants as likely to default or not. I also use balancing and sampling techniques to measure the impacts on the models' abilities to detect the minority class (positive default classifications). Results: Among the models used, the true positive rate, false positive rate, and positive predictive value are greater for men than women. This is true both when gender is included and excluded as a feature. We also see that across models, removing gender as a feature in the training process increases the discrepancy between men and women with respect to the positive predictive value. However, across models, removing gender as a feature decreases the discrepancy between men and women with respect to the true positive rate and the false positive rate. Further, we see that balancing gender in the training data has mixed effects.