[論文レビュー] Dissecting Performative Prediction: A Comprehensive Survey

Thomas Kehrenberg, Javier Sanguino+2 more|arXiv (Cornell University)|Feb 10, 2026

Generative Adversarial Networks and Image Synthesis被引用数 0

ひとこと要約

この調査はパフォーマティブ予測を定義し、安定性と最適性を形式化し、情報アクセスに基づく分布マップを分類し、分布マップの実装と最適化手法を調査し、関連分野とのつながりを強調して今後の研究を促す。

ABSTRACT

The field of performative prediction had its beginnings in 2020 with the seminal paper "Performative Prediction" by Perdomo et al., which established a novel machine learning setup where the deployment of a predictive model causes a distribution shift in the environment, which in turn causes a mismatch between the distribution expected by the predictive model and the real distribution. This shift is defined by a so-called distribution map. In the half-decade since, a literature has emerged which has, among other things, introduced new solution concepts to the original setup, extended the setup, offered new theoretical analyses, and examined the intersection of performative prediction and other established fields. In this survey, we first lay out the performative prediction setting and explain the different optimization targets: performative stability and performative optimality. We introduce a new way of classifying different performative prediction settings, based on how much information is available about the distribution map. We survey existing implementations of distribution maps and existing methods to address the problem of performative prediction, while examining different ways to categorize them. Finally, we point out known and previously unknown connections that can be drawn to other fields, in the hopes of stimulating future research.

研究の動機と目的

• Define performative prediction and emphasize distribution shift caused by model deployment.
• Introduce a classification of distribution maps based on information access.
• Survey existing distribution-map implementations and methods to address performativity.
• Discuss optimization approaches for achieving stable and optimal models.
• Highlight connections between performative prediction and related fields to inspire future research.

提案手法

• Formalize risk using the distribution map D(·) and define performative risk PR(θ)=Risk(θ, D(θ)).
• Define performative stability (PS) as a fixed-point minimizer on its induced distribution.
• Define performative optimality (PO) as the minimizer of Risk(θ, D(θ)) over θ.
• Classify distribution maps by information availability and access to the environment’s response.
• Survey algorithms and methods to find stable points and performative-optimal points.
• Discuss stateful variants via transition maps and connections to MDPs/Bandits.

実験結果

リサーチクエスチョン

• RQ1 What are the precise definitions of performative stability and performative optimality?
• RQ2 How can distribution maps be classified based on the amount of information and access about environmental response?
• RQ3 What algorithms exist to compute stable points and performative-optimal points, and under what conditions do they converge?
• RQ4 How does performative prediction relate to and differ from Stackelberg games, bandits, and reinforcement learning?
• RQ5 What are the implications of stateful (transition-map) extensions for gradual adaptation?

主な発見

• Provide a formal framework for performative prediction with a distribution map D(·) and performative risk PR(θ).
• Distinguish between stable points (PS) and optimal points (PO), noting they are not generally aligned.
• Offer a novel classification of performative settings by distribution-map access and information about environmental response.
• Survey methods to estimate and utilize distribution maps, and optimization approaches for PS and PO.
• Explore stateful extensions (transition maps) and relate PP to MDPs, bandits, and Stackelberg games.
• Highlight connections to related fields (adversarial settings, algorithmic recourse, delayed impact of fairness) to stimulate cross-disciplinary work.