Active learning using blending query strategy with multi-armed bandit on imbalanced data classification

Abstract

This thesis considers a query strategy in active learning referred to as blending query strategy to gain robustness to data imbalance in a classification task. Balancing between variance and bias of the estimated decision boundary at each step of the active learning framework requires a strategy to sample the most difficult data samples as well as preferential selection of samples in the minority class. To achieve this feat, a multi-armed bandit is incorporated in blending the following two strategies: (1) the uncertainty sampling selecting the most uncertain samples and (2) the minority preferential query strategy selecting the most informative minority class samples. Blending of the two strategies at each iteration is conducted based on a probability defined as a function of validation error and sample uncertainty. The uncertainty of a sample is measured by the Bayesian Active Learning by Disagreement (BALD) which requires evaluation of sample entropy. Here the entropy is obtained using a Bayesian deep neural network. Experimental results on binary data sets(HIVA, ZEBRA, Spambase) as well as multi-class data set(MNIST) show that the performance of the proposed blending query strategy outperform other state-of-the-art algorithms such as Intra-Class Clustering (ICC), and Co-selecting algorithm. Also, various aspects are studied for the extension of the proposed algorithm.