Bilevel Optimization with Coupled Decision-Dependent Distributions

Authors: Songtao Lu

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our theoretical analysis is corroborated through a series of numerical experiments, wherein we evaluate the performance of the bilevel performative prediction algorithms alongside non-performative counterparts in the context of meta strategic learning problems. To validate our theoretical results and demonstrate the stability of Bi-SGD under data distribution shifts, we conduct experiments on both synthetic and real data sets. The results underscore the efficacy of Bi-RRM and Bi-SGD.
Researcher Affiliation Industry Songtao Lu 1 1IBM Research, Thomas J. Watson Research Center, Yorktown Heights, New York 10598, USA. Correspondence to: Songtao Lu <songtao@ibm.com>.
Pseudocode No The paper describes the algorithms (Bi-RRM and Bi-SGD) using mathematical equations and textual descriptions, but does not provide a formally labeled 'Algorithm' or 'Pseudocode' block.
Open Source Code No The paper does not include an explicit statement about releasing open-source code or a link to a code repository.
Open Datasets Yes In the numerical experiments, we employ the UCI machine learning repository spambase data set (Hopkins et al., 1999) for binary classification. We also perform the numerical experiments using the UCI sentiment labeled sentences data set (Kotzias et al., 2015), specifically the Amazon reviews subset, which is a prominent area of research in natural language processing.
Dataset Splits No The paper describes meta-training and meta-testing sets and their use (e.g., '5 samples used for meta-training and the remaining samples used for meta-testing'), but does not explicitly define a separate 'validation' split with specific percentages or counts for hyperparameter tuning or early stopping. The term 'validation' is not used in the context of data splits.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory used for running the experiments.
Software Dependencies No To vectorize the sentences, we utilize the Count Vectorizer from the scikit-learn library, generating word count vectors for each sentence. However, no specific version number for scikit-learn or any other software dependency is provided.
Experiment Setup Yes In our numerical experiments, we set the problem dimension as 5 and the total number of data samples as 50. The parameter ε represents both εx and εy. We choose λx = λy = 1 10 3, the minibatch size as 5, and use the same step size 1/ r for both αr and βr in Bi-SGD. We set the values of λx, λy to 1 10 3 and 1 respectively, and the step size is chosen as 0.5/ 10 + r for both Bi-SGD and SGD.