We propose a model of rational preference change that is identifiable, empirically testable and founded on two normative principles. First, the decision maker (DM) must be able to justify her preference change by making attributes of the alternatives relevant or irrelevant. For instance, an employer who stops discriminating against candidates based on race or gender must make the attribute "race" or "gender" irrelevant for all her future choices. Second, DM must be consistent in her preference change. For instance, in the future the employer cannot make relevant again the attribute "race" or "gender", unless she makes (ir)relevant other complementary attributes. We show that two axioms on choice data which reflect these principles are necessary and sufficient to represent preference change by the maximization of a meta-preference. Finally, we apply our model to derive new insights on the polarization of political preferences.