• Determined that briefly summarising the current state of algorithmic recourse and performative prediction is a necessary first step
• Discussed the original credit-loan setup from Perdomo et al., 2020.
  • it seems potentially unrealistic that the labels don't change
• Building a realistic simulation suite for performative prediction / (dynamic) algorithmic recourse might be one avenue for the project to take. This was motivated by the deficiencies brought up in the above-referenced discussion.
• Should the setup allow black box models?
• Should the setup allow for asynchronicity?
• Should the setup involve competition between the players/clients?
• Should the training distribution be fully observed?
• Our goal is, I think, to try to bridge the gap between the existing PP literature, which is largely theoretical/toy, and real-world realisations of the dynamic problems it represents. This begins with developing some certifiably-realistic testbed, in the vein of WILDS.

typst project: https://typst.app/project/wHE0EGvje8APHT1pdoj-bm

Since we still need to figure out our exact angle of attack, getting a solid grasp of the current state of Performative Prediction and Algorithmic Recourse seems like a necessary preliminary step. Let's start by having a page or so summarising such.

On the AF side, this 2022 survey coauthored by Scholkopf looks quite excellent.

Some PP/SC papers that may be interesting/relevant:

• Multiplayer Performative Prediction: Learning in Decision-Dependent Games -- excellent game-theoretic perspective on the problem that considers not only the performative aspect but also competition between players (which factors into the distribution map and loss function (if non-separable)). The semi-synthetic ride-share dataset/simulation could certainly be relevant to us. The much-abbreviated version of this paper was published at ICML.
• Strategic Classification in the Dark -- addresses one of the deficiencies of the PP setup raised during the meeting, namely classifier being fully revealed (observable) to the agents/clients/players. Abstract excerpt: "...in many real-life scenarios of high-stake classification (e.g., credit scoring), the classifier is not revealed to the agents, which leads agents to attempt to learn the classifier and game it too. In this paper we generalize the strategic classification model to such scenarios. We define the price of opacity as the difference in prediction error between opaque and transparent strategy-robust classifiers, characterize it, and give a sufficient condition for this price to be strictly positive, in which case transparency is the recommended policy....