Minimax Risk Classifiers with 0-1 Loss

Authors: Santiago Mazuelas, Mauricio Romero, Peter Grunwald

JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This section shows four sets of numerical results that describe how the proposed techniques can enable to learn MRCs efficiently and to obtain reliable and tight performance bounds at learning. We utilize 12 common datasets from the UCI repository (Dua and Graff, 2017) with characteristics given in Table 1.
Researcher Affiliation Academia Santiago Mazuelas EMAIL Basque Center for Applied Mathematics (BCAM) IKERBASQUE-Basque Foundation for Science Bilbao 48009, Spain Mauricio Romero EMAIL Federal University of Bahia Ondina 40170, Brazil Peter Gr unwald EMAIL National Research Institute for Mathematics and Computer Science (CWI) Amsterdam 94079, Netherlands
Pseudocode Yes Algorithm 1 ASM for MRC learning... Algorithm 2 Efficient ASM for MRC learning
Open Source Code Yes MRCs implementation is available in the open-source Python library MRCpy (Bondugula et al., 2023) https://Machine Learning BCAM.github. io/MRCpy/.
Open Datasets Yes We utilize 12 common datasets from the UCI repository (Dua and Graff, 2017) with characteristics given in Table 1.
Dataset Splits Yes In order to reproduce the ideal case, in these numerical results we use Adult and Pulsar datasets, the mean τ is calculated using all the samples while 1, 000 samples are randomly sampled for training in 50 repetitions... In particular, for each value of λ0 and ρ we averaged over 10-fold stratified partitions the error and bounds of MRCs and DRLRs... Specifically, we generate 20 random stratified splits with 20% test samples.
Hardware Specification No No specific hardware details (like GPU/CPU models, memory amounts, or detailed computer specifications) are mentioned in the paper.
Software Dependencies No The paper mentions "open-source Python library MRCpy" but does not specify Python version or any other library versions used.
Experiment Setup Yes In all the numerical experiments, if not stated otherwise, we take λ0 = 0.3 and use D = 500 random Fourier features corresponding with the scaling parameter σ = p d/2 for d the number of instances components... SVM-CV selects the scaling parameter for which a SVM classifier achieves the smallest cross-validation error over 10-fold partitions of the training data... For all datasets, MRCs utilize confidence vectors as in (35) with λ0 = 0.3 and DRLRs utilize Wasserstein radious of ρ = 0.003 as in Shafieezadeh-Abadeh et al. (2015).