Online Multi-threshold Learning with Imbalanced Data Stream

Xufen Cai, Min Yang, Rong Zhu, Xiaoyan Li, Long Ye, Qin Zhang

Abstract: This paper addresses the imbalanced data problem in an online fashion based on multi-threshold learning. The majority of existing works on processing large-scale imbalanced data stream assume a prior distribution of data based on a training dataset, while we consider a more challenging setting without any assumption of the prior, and propose an online multi-threshold learning (OMTL) method by simultaneously learning multiple classifiers with different threshold based on F-measure incremental updating. The proposed approach shows its potentials on recent benchmark datasets compared to previous cost-sensitive and threshold fine-tuning based online learning algorithms.