Leveraging Constraints for User-Centric Selection of Predictive Features

Feature selection identifies the most valuable predictors in a dataset. Thus, feature-selection techniques are popular for obtaining small, interpretable, yet highly accurate prediction models. While optimizing technical quality metrics, standard feature-selection techniques might not satisfy user needs for two reasons. First, existing methods do not consider domain knowledge. Such domain knowledge can restrict which feature combinations make sense to users. Second, traditional feature-selection techniques typically yield only one feature set, which might not suffice in some scenarios. For example, users might be interested in finding different feature sets with similar prediction quality, offering alternative explanations of the data.
Constraints on feature sets alleviate both these shortcomings. First, constraints allow users to express domain knowledge, e.g., known physical laws, novel scientific hypotheses, etc. Second, constraints can formalize the notion of alternative feature sets.
Our research studies different types of constraints that make feature selection more user-centric. We investigate how to formulate and integrate such constraints into existing feature-selection techniques. ... mehrFurther, we study the impact of constraints on feature-selection results, e.g., if prediction quality remains stable under constraints. Our experiments show that it often is possible to find high-quality feature sets adhering to user constraints.