RMI-RRG: A Soft Protocol to Postulate Monotonicity Constraints for Tabular Datasets

Iko Vloothuis, Wouter Duivesteijn

Abstract: Ensuring that a predictive model respects monotonicity constraints can enhance societal acceptance of such models. Literature on monotone classification shows that it can even improve classifier performance. However, a set of applicable monotonicity constraints is often assumed as input for the model. We propose RMI-RRG: a soft protocol that can be employed to postulate monotonicity constraints for any tabular dataset. The protocol encompasses consensus from scientific literature, aggregating the strength of (anti-)monotonicity relations in an RMI Table, aggregating the effect of imposing more constraints on the number of relabelings required to fully monotonize the dataset in a Required Relabelings Graph (RRG), and inspecting the effect on the comparability rate. We illustrate the deployment of the protocol on six datasets, arriving at some conclusions that deviate from conclusions from (mutually disagreeing) existing literature, and showing how individual steps in the protocol each have their role to play in arriving at a final postulate.