Revisiting the Predictability of Performative, Social Events
Authors: Juan Carlos Perdomo
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | In this work, we revisit this question using modern mathematical tools. Framing our analysis in the language of performative prediction a recent learning-theoretic framework introduced by Perdomo et al. (2020) that formalizes the causal aspects of social prediction we establish the following result. One can always efficiently find predictions that actively shape the data distribution over outcomes and simultaneously satisfy rigorous validity guarantees like multi-calibration (H ebert-Johnson et al., 2018) or outcome indistinguishability (Dwork et al., 2021). Moreover, the statistical and computational complexity of finding these predictors is, in many cases, just as easy as in supervised learning, where the distribution is static. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Harvard University, Boston MA, USA. Correspondence to: Juan Carlos Perdomo <EMAIL>. |
| Pseudocode | Yes | A. The K29 Algorithm Input: A kernel k : (X [0, 1])2 R For t = 1, 2, . . . 1. Given history {(xi, pi, yi)}t 1 i=1 and current features xt define St : [0, 1] R as St(p) = P t 1 i=1 k((xt, p), (xi, pi))(yi pi). 2. If St(1) 0 predict pt = 1. Else if St(0) 0 predict pt = 0. 3. Else, run binary search to return pt [0, 1] such that St(pt) = 0 Figure 2. At every round, the algorithm implicitly publishes a function ft : X [0, 1] where predictions pt = f(xt) are chosen by solving a simple optimization problem defined with respect to the history {(xi, pi, yi)}t 1 i=1. |
| Open Source Code | No | No explicit statement about releasing code or providing a link to a repository is found in the paper. |
| Open Datasets | No | The paper is theoretical and focuses on mathematical proofs and algorithmic analysis rather than empirical evaluation on specific datasets. It does not mention or provide access to any publicly available datasets. |
| Dataset Splits | No | No specific datasets are used for empirical evaluation, hence no dataset split information is provided. |
| Hardware Specification | No | The paper is theoretical and does not describe any experiments that would require specific hardware. No hardware specifications are provided. |
| Software Dependencies | No | The paper is theoretical and does not describe any experiments that would require specific software dependencies with version numbers. No such details are provided. |
| Experiment Setup | No | The paper is theoretical and does not describe any experimental setup or training configurations, including hyperparameters. |