Human-in-the-Loop Feature Selection

Feature selection is a crucial step in the conception of Ma-chine Learning models, which is often performed via data-driven approaches that overlook the possibility of tappinginto the human decision-making of the model's designers andusers. We present ahuman-in-the-loopframework that inter-acts with domain experts by collecting their feedback regard-ing the variables (of few samples) they evaluate as the mostrelevant for the task at hand. Such information can be mod-eled via Reinforcement Learning to derive a per-example fea-ture selection method that tries to minimize the model's lossfunction by focusing on the most pertinent variables from ahuman perspective. We report results on a proof-of-conceptimage classification dataset and on a real-world risk classi-fication task in which the model successfully incorporatedfeedback from experts to improve its accuracy.