Learning to Incentivize Improvements from Strategic Agents
Abstract: Machine learning systems are often used in settings where individuals adapt their features to obtain a desired outcome. In such settings, strategic behavior leads to a sharp loss in model performance in deployment. In this work, we aim to address this problem by learning classifiers that encourage decision subjects to change their features in a way that leads to improvement in both predicted and true outcome. We frame the dynamics of prediction and adaptation as a two-stage game, and characterize optimal strategies for the model designer and its decision subjects. In benchmarks on simulated and real-world datasets, we find that classifiers trained using our method maintain the accuracy of existing approaches while inducing higher levels of improvement and less manipulation.
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: Here is a summary of the major changes we made in our revision: * Reviewer 5yXo (colored blue): 1) Clarify the assumption we make in Proposition 6, which does not correspond to the covariate shift setting. 2) Clarify the assumption we make in our experimental section on the german dataset. 3) Add a broader impact section at the end of the paper. * Reviewer EU7Q (colored pink): 1) Add a detailed description of the Stackelberg game, including the two parties, action spaces as well as pointers to more detailed discussions for their corresponding utility function in Section 2.1. 2) Add descriptions of the ideal utility function for the decision-maker in Section 2.3. 3) Fix the reference formatting problem. 4) Clarify the ambiguity for true qualification function $y(x)$. 5) Add a limitation on assuming the separation of the causal and non-causal features. * Reviewer TLHt (colored red): 1) Add the GitHub repository link for the code. 2) Add the proof for Proposition 3. 3) Add a clarification on the subscript and superscript notation. 4) Add a discussion on the concept of actionability and the concept of target trail in causal intervention.
Supplementary Material: zip
Assigned Action Editor: ~Pascal_Poupart2
Submission Number: 768