Ensemble model that minimizes the misclassification cost for imbalanced class credit data and its explanation using LIME.

Recently ensemble models have been proposed for credit scoring applications to alleviate the negative impact of imbalanced class data that can cause deterioration in prediction performance. Ensemble model have shown promising results in terms of model accuracy in previous studies. However, rather than model accuracy, the ultimate interest in credit scoring is to minimize the financial losses that can be more precisely quantified using total misclassification cost (TMC). Despite TMC is thought to be an effective indicator to assess the financial losses in credit scoring, unfortunately, TMC has yet been adequately studied for imbalanced credit data. This paper investigates the performance of several representative ensemble models over the imbalanced data using TMC as a model assessment tool. In addition, due to the fact that the bank is highly conservative in using new credit scoring models, we employ the LIME technique to interpret the best-selected ensemble model and provide some understanding of the predicted results. The finding shows that the ensemble model, particularly the bagging model, can provide a minimum TMC and higher discriminality power for imbalanced class data compared to the benchmark models (i.e. single classifier). Furthermore, LIME offers a reasonable interpretation of the prediction based on higher R² value produced. [ABSTRACT FROM AUTHOR]