Often in learning situations, the exceptional class that one seeks to learn forms a small proportion of the total, so that a conventional learning system that minimizes the error can easily fail to recognize the exceptional elements.

One way to handle this issue is to increase the cost of misclassifying the rare examples. The other major issue addressed in this paper is that of learning from an online data stream. While recognizing that these two problems have been well studied in the past, the research into the combined problems is more limited.

The authors review a number of boosting and bagging algorithms, giving explicit descriptions of the algorithms. They follow this review by presenting online versions of the algorithms that are suitable for learning online in cases where the exceptional class is relatively infrequent. Once again, each algorithm is explicitly described, although the reader will benefit from consulting earlier descriptions of online bagging to best understand the descriptions. The authors describe detailed experiments that compare their online learning algorithms with the batch counterparts. They also discuss whether the results of the online algorithms converge to those of their batch mode counterparts. In some cases they do, but in others they cannot be expected to.

The paper can be viewed not only as giving interesting results, but also as providing a useful survey of its field.