Class imbalances versus small disjuncts

Abstract

It is often assumed that class imbalances are responsible for significant losses of performance in standard classifiers. The purpose of this paper is to the question whether class imbalances are truly responsible for this degradation or whether it can be explained in some other way. Our experiments suggest that the problem is not directly caused by class imbalances, but rather, that class imbalances may yield small disjuncts which, in turn, will cause degradation. We argue that, in order to improve classifier performance, it may, then, be more useful to focus on the small disjuncts problem than it is to focus on the class imbalance problem. We experiment with a method that takes the small disjunct problem into consideration, and show that, indeed, it yields a performance superior to the performance obtained using standard or advanced solutions to the class imbalance problem.

Overlapping, Rare Examples and Class Decomposition in Learning Classifiers from Imbalanced Data

“…Two imbalance ratios and five different cardinalities of data sets are reported in columns. in [30]. For the number of examples greater than 400 examples there was no significant influence of decreasing the imbalance ratio.…”

“…target concepts (classes) were decomposed into sub-concepts [30]. To check the influence of increasing the level of decomposition she and her co-authors carried our many experiments with artificially generated data.…”

“…it contains very small subclusters". On the other hand, this also means that in much larger data where subclusters could be represented by a reasonable number of examples, the imbalance ratio alone will not decrease so much the classification performance [30].…”

“…As a practical consequence special approaches to handle the problem of small disjuncts of imbalanced concepts were proposed in [29,30]. They are based on specialized over-sampling of minority class, i.e.…”

“…Recent works also suggest that there are other factors that contribute to difficulties. The most well known studies with artificial data are the works of Japkowicz [29,30], who showed that simple class imbalance ratio was not the main difficulty. The degradation of performance was also related to other factors, mainly to decomposition of the minority class into many sub-clusters with very few examples.…”