<ul><li>Machine learning is widely used in societal decision-making and it is important to consider how classified agents will react to learning algorithms.</li><li>Recent research highlights properties of learnability when agents genuinely improve in order to achieve desirable classifications.</li><li>This paper characterizes learnability with improvements across various aspects and introduces an asymmetric variant of minimally consistent concept classes.</li><li>The study provides insights into learning with improvements under different settings, achieving lower generalization error and resolving open questions in the field.</li></ul>

Conservative classifiers do consistently well with improving agents: characterizing statistical and online learning

Discover more