Countering Racial Discrimination in Algorithmic Lending: A Case for Model-Agnostic Interpretation Methods