Performative Prediction

Authors: Juan Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, Moritz Hardt

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We next examine the convergence of repeated risk minimization and repeated gradient descent in a simulated strategic classification setting. We run experiments on a dynamic credit scoring simulator in which an institution classifies the creditworthiness of loan applicants. To run our simulations, we construct a distribution map D(θ), as described in Figure 2. For the base distribution D, we use a class-balanced subset of a Kaggle credit scoring dataset (Kaggle, 2012). Figures 1 and 3 show experimental results.
Researcher Affiliation Collaboration 1University of California, Berkeley 2MH is a paid consultant for Twitter.
Pseudocode Yes Figure 2: Distribution map for strategic classification. Input: base distribution D, classifier fθ, cost function c and utility function u Sampling procedure for D(θ): 1. Sample (x, y) D 2. Compute x BR arg maxx0 u(x0, θ) c(x , x) 3. Output sample (x BR, y)
Open Source Code No The paper does not provide an explicit statement about the release of its source code or a link to a code repository.
Open Datasets Yes For the base distribution D, we use a class-balanced subset of a Kaggle credit scoring dataset (Kaggle, 2012).
Dataset Splits No The paper states it uses a "class-balanced subset of a Kaggle credit scoring dataset" but does not provide specific details on how the dataset was split into training, validation, or test sets, such as percentages, counts, or a description of cross-validation.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory, or cloud instances) used to run the experiments.
Software Dependencies No The paper mentions using a 'logistic regression classifier' but does not specify any software dependencies with version numbers, such as programming languages, libraries, or frameworks used for implementation.
Experiment Setup Yes Further details about the ex perimental setup may be found in Appendix G. In Appendix G, it states: "The weights θ are initialized to zero." "We use a logistic loss function augmented with a strongly convex regularization term of the form λ kθk22 . We used λ = 10 5 ." "We ran each experiment for 1000 iterations." "For the repeated gradient descent procedure, we used a step size η = 0.001."