Analyzing Adaptive and Non-Adaptive Online Learners on Imbalanced Evolving Streams

Abstract

Recently, the combined problem of online class imbalance and concept drift (OCI-CD) has received much interest. The effect of this combined problem on the state-of-the-art of online adaptive and non-adaptive learners is not widely investigated. This work explores the impact of the parameters such as current imbalance ratio, length of the stream, type of the drift, levels of the drift, and state of the imbalance (static or dynamic) on adaptive and non-adaptive online learners. The experimental results demonstrate that each parameter considered for the study has a notable impact on learner performance: (a) the minority class performance decreases with the increase of the degree of imbalance, (b) non-adaptive learners are much prone to class imbalance, concept drift and the combined problem of both drifts than adaptive learners, (c) adaptive learners are only susceptible to class imbalance drifts, (d) the impact of the dynamic degree of imbalance on learner performance is more adverse than the static degree of imbalance and (e) the adaptive large scale support vector machine yields stable performance to all the parameters considered for the study. Further, directions to develop new approaches are also presented based on these findings.