Strategic classification studies the interaction between a learning model and agents who proactively modify their features to obtain more favorable classifications.
Traditionally, these models assume that agents are perfectly rational utility maximizers.
We challenge this assumption by incorporating Prospect Theory, under which agents evaluate outcomes as gains or losses relative to a reference point rather than solely in absolute terms.
We study two sources of these reference points. First, we consider exogenous reference points, determined independently of the system, and show that high reference points can improve classification accuracy by inducing agents to exert greater effort, but at a cost to the users.
Second, we consider system-influenced reference points, where the system can shape agents’ expectations, and analyze how this additional degree of control affects strategic behavior, predictive performance, and user welfare.