Adversarial Kernel Sampling on Class-imbalanced Data Streams

Abstract

This paper investigates online active learning in the setting of class-imbalanced data streams, where labels are allowed to be queried of with limited budgets. In this setup, conventional learning would be biased towards majority classes and consequently harm the performance. To address this issue, imbalance learning technique adopts both asymmetric losses and asymmetric queries to tackle the imbalance. Although this approach is effective, it may not guarantee the performance in an adversarial setting where the actual labels are unknown, and they may be chosen by the adversary To learn a promising hypothesis in class-imbalanced and adversarial environment, we propose an asymmetric min-max optimization framework for online classification. The derived algorithm can track the imbalance and bound the choices of an adversary simultaneously. Despite the promising result, this algorithm assumes that the label is provided for every input, while label is scare and labeling is expensive in real-world application. To this end, we design a confidence-based sampling strategy to query the informative labels within a budget. We theoretically analyze this algorithm in terms of mistake bound, and two asymmetric measures. Empirically, we evaluate the algorithms on multiple real-world imbalanced tasks. Promising results could be achieved on various application domains.