Search papers, labs, and topics across Lattice.
This paper introduces a two-stage robust optimization framework that accounts for decision-dependent cost uncertainty in strategic classification, addressing a critical gap in existing models that treat manipulation costs as fixed. By recognizing that the costs of strategic behavior evolve based on past algorithmic decisions, the authors demonstrate that their approach not only reduces uncertainty but also effectively mitigates the tendency for users to game the system. The key finding reveals that incorporating policy-dependent costs leads to improved long-term outcomes in algorithmic decision-making environments.
Awareness of evolving manipulation costs can significantly enhance the robustness of algorithmic systems against strategic gaming.
Humans facing algorithmic decision systems have been found to ``game'' them by altering their input data (at a cost to them) in order to favorably change the algorithmic outcomes they receive (at a cost to the algorithm). The growing literature on strategic classification seeks to develop robust machine learning algorithms that account for, and reduce, unwanted strategic behavior. A limitation of these existing works is that they assume the cost of strategic behavior to be fixed and independent of the classifier's decision. In practice, however, manipulation costs evolve and depend on past algorithmic decisions: today's decisions influence tomorrow's costs. This paper proposes and analyzes a two-stage robust optimization framework with a decision-dependent uncertainty set to capture such dependencies. We highlight that awareness of policy-dependent costs not only reduces uncertainty, but also better curtails gaming of the algorithmic system over time.
