Class-Weighted Classification: Trade-offs and Robust Approaches

Authors: Ziyu Xu, Chen Dan, Justin Khim, Pradeep Ravikumar

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we empirically demonstrate the efficacy of LCVa R and LHCVa R on improving class conditional risks.
Researcher Affiliation Academia 1Machine Learning Department, Carnegie Mellon University, Pennsylvania, United States 2Computer Science Department, Carnegie Mellon University, Pennsylvania, United States.
Pseudocode No No explicitly labeled pseudocode or algorithm block is present in the main text provided. The paper mentions pseudocode might be in the appendix, which is not available.
Open Source Code Yes Code for reproducing the results in this section can be found at https://www.github.com/neilzxu/ robust_weighted_classification.
Open Datasets Yes Real World Datasets We also experiment on the Covertype dataset taken from the UCI dataset repository (Dua & Graff, 2017).
Dataset Splits No The paper mentions '10,000 data points for both train and test sets' for synthetic data and a 'train-test split' for Covertype, but does not specify a validation set or explicit train/validation/test split percentages.
Hardware Specification No No specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running experiments are provided in the paper.
Software Dependencies No The paper mentions training a logistic regression model with gradient descent on a cross entropy loss, but does not provide specific software dependencies with version numbers (e.g., 'Python 3.8, PyTorch 1.9').
Experiment Setup Yes LCVa R The empirical formulation optimizes the dual formulation, in which α is a hyperparameter: \LCVa Rα(f) = min λ R i=1 bpi( b Ri(f) λ)+ + λ. To reduce the number of hyperparameters to only c (0, 1] and κ (0, ), we calculate αi as follows: α(κ,c) i = c bpi 1/κ Pk j=1 bpj 1/κ. We train a logistic regression model with gradient descent on a cross entropy loss, which acts as a convex surrogate loss for zero-one risk.