The authors performed a comprehensive study on the relationship between accuracy and diversity using an ensemble of ARTMAP neural network classifiers tested with five datasets from the University of California, Irvine (UCI) repository. Their results indicate that lower accuracies (evaluated using a novel method of rankings) are obtained when using homogenous ensembles as opposed to heterogeneous ones--which is expected since more diverse ensembles will give better performance. The results also indicate that using the RePART neural network in the ensemble improves accuracy in almost all of the cases.

A major finding of their study is that the correlation between diversity and accuracy is stronger when using a smaller number of classifiers. Hence, diversity may not be a useful measure for deciding the classifiers to be used in an ensemble when many classifiers are used.

There is one minor issue: the authors frequently denote ARTMAP neural networks to be faster than most other neural networks. While this is true during the training step, most neural networks are comparatively fast during evaluation and testing.

In general, the paper is well written and would be useful to anyone interested in using ARTMAP neural networks (there is a good introduction on ARTMAP networks and recent improved variants such as ARTMAP-IC and RePART). It will also prove to be useful to researchers when determining the types of neural networks (or any type of classifier, for that matter) to use when employing an ensemble of classifiers to obtain improved performance.