This repository contains source files of the paper titled Bandit Multiclass Linear Classification: Efficient Algorithms for the Separable Case authored by Alina Beygelzimer, Dávid Pál, Balázs Szörényi, Devanathan Thiruvenkatachari, Chen-Yu Wei and Chicheng Zhang. The paper was accepted at ICML 2019.
Bandit multiclass classification is a problem where in each round an algorithm receives a feature vector and it needs to predict one of K classes, after which it receives binary feedback whether or not the class was correct. We design efficient algorithms under the assumption that data are linearly separable.
The source codes for the experiments are provided in another repository: https://github.com/bahh723/separable-bandit-classification.git