Rashomon Capacity: A Metric for Predictive Multiplicity in Classification

Authors: Hsiang Hsu, Flavio Calmon

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our numerical experiments illustrate how Rashomon Capacity captures predictive multiplicity in various datasets and learning models, including neural networks.
Researcher Affiliation Academia Hsiang Hsu and Flavio P. Calmon John A. Paulson School of Engineering and Applied Sciences, Harvard University hsianghsu@g.harard.edu, flavio@seas.harvard.edu
Pseudocode Yes We describe next a method (described in detail in Algorithm SM. 2) based on weight perturbation that obtains c models in the Rashomon subset for each sample.
Open Source Code Yes Code to reproduce our experiments is available at https://github.com/HsiangHsu/rashomon-capacity.
Open Datasets Yes We illustrate how to measure, report, and potentially resolve predictive multiplicity of probabilistic classifiers using Rashomon Capacity on UCI Adult [25], COMPAS [26], HSLS [27], and CIFAR-10 datasets [8].
Dataset Splits No The paper mentions 'Each point is generated with 5 repeated splits of the dataset' and refers to 'training details' in the Supplementary Materials, but does not provide specific percentages or absolute counts for train/validation/test splits in the main text.
Hardware Specification No The paper states 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] Provided in the SM.' However, specific hardware details like GPU/CPU models are not provided in the main text.
Software Dependencies No The paper states 'Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] All provided in the SM.' However, specific software dependencies with version numbers are not provided in the main text.
Experiment Setup No The paper states 'For more information on the datasets, neural network architectures, and training details, see Section SM. 3.2.' It defers detailed experimental setup, including hyperparameters, to the Supplementary Materials and does not provide them in the main text.