Predictive models with end user preference

Abstract

AbstractClassical machine learning models typically try to optimize the model based on the most discriminatory features of the data; however, they do not usually account for end user preferences. In certain applications, this can be a serious issue as models not aware of user preferences could become costly, untrustworthy, or privacy‐intrusive to use, thus becoming irrelevant and/or uninterpretable. Ideally, end users with domain knowledge could propose preferable features that the predictive model could then take into account. In this paper, we propose a generic modeling method that respects end user preferences via a relative ranking system to express multi‐criteria preferences and a regularization term in the model's objective function to incorporate the ranked preferences. In a more generic perspective, this method is able to plug user preferences into existing predictive models without creating completely new ones. We implement this method in the context of decision trees and are able to achieve a comparable classification accuracy while reducing the use of undesirable features.