Weighting machine learning solutions by economic and institutional context for decision making

Abstract

[EN] It is quite common that machine learning approaches reach high accuracy forecast rates in imbalanced datasets. However, the results in the category with few instances are usually low. This paper seeks to improve the results obtained applying different techniques (such as bagging, boosting or random forests) with the inclusion of cost matrices. We propose applying the actual costs incurred by the company for misclassification of instances as a cost matrix. This approach, along with an economic analysis of the different solutions, makes it possible to incorporate a business perspective in the decision making process. The approach is tested on a publicly available dataset. In our example, the best ratings are obtained by combining the cost matrix with random forests. However, our analysis shows that the best technical solution is not always the best economical solution available. A company cannot always implement the optimal solution, but has to adopt a solution constrained by its social, institutional and economic context. Once an economic analysis is carried out, it seems the final decision of the company will depend on its economic situation and its institutional policy.