Classification with Valid and Adaptive Coverage

Authors: Yaniv Romano, Matteo Sesia, Emmanuel Candes

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on synthetic and real data demonstrate the practical value of our theoretical guarantees, as well as the statistical advantages of the proposed methods over the existing alternatives.
Researcher Affiliation Academia Yaniv Romano Department of Statistics Stanford University Stanford, CA, USA yromano@stanford.edu Matteo Sesia Department of Data Sciences and Operations University of Southern California Los Angeles, CA, USA sesia@marshall.usc.edu Emmanuel J. Candès Departments of Mathematics and of Statistics Stanford University Stanford, CA, USA candes@stanford.edu
Pseudocode Yes Algorithm 1: Adaptive classification with split-conformal calibration
Open Source Code Yes The Python package at https://github.com/msesia/arc implements our methods. This repository also contains code to reproduce our experiments.
Open Datasets Yes The methods are tested on two well-known data sets: the Mice Protein Expression data set3 and the MNIST handwritten digit data set. (Footnote 3: https://archive.ics.uci.edu/ml/datasets/Mice+Protein+Expression)
Dataset Splits Yes Algorithm 1: Input: data {(Xi, Yi)}n i=1... Randomly split the training data into 2 subsets, I1, I2. ... Algorithm 2: Input: data {(Xi, Yi)}n i=1... Randomly split the training data into K disjoint subsets, I1, . . . , IK, each of size n/K.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No We compare the performances of Algorithms 1 (SC) and 2 (CV+, JK+)... We explore 3 different black-boxes: an oracle... a support vector classifier (SVC) implemented by the sklearn Python package; and a random forest classifier (RFC) also implemented by sklearn...
Experiment Setup Yes We fix α = 0.1 and assess performance in terms of marginal coverage, conditional coverage, and mean cardinality of the prediction sets.