Making Lemonade Out Of LIME: A Comparative Analysis Of Interpretable Machine Learning Methods

Abstract

Increasingly complex machine learning methods have been proposed to enhance the capabilities of algorithmic decision making. While more intricate black box models often result in more accurate classification, there is a trade-off between complexity and interpretability. Model interpretation is important to practitioners hoping to improve their models, end users seeking trust in the predictions that models make, and regulatory agencies seeking to audit complex algorithms. Here, we apply two proposed solutions to model interpretation, LIME and decision sets, to a loan classification problem using data from the LendingClub, which consists of 42,538 samples and 145 features. LIME offers post-hoc interpretation of individual predictions while decision sets is a machine learning algorithm that optimizes over both accuracy and interpretability. Our results suggest that LIME offers flexibility, as it can be easily adapted to explain different models but is not well suited for use cases where features are numerous and highly correlated. The decision set algorithm on the other hand, presents a highly interpretable classifier in the form of 28 rules, which classifies loan outcomes with 75.75% accuracy compared to 85.79% accuracy for a decision tree trained on the same data.