In many classification problems, e.g., text classification, with tthe increase of class number, the discrepancy between different classes would reduce. So it would become more difficult to perform dimension reduction and classification techniques effectively.
Recently, We have used the 20Newsgroup to create 8 datasets with varying class number to examine the effect of varying class number on the performance of various classifiers.
The datasets were used in our recent DMKD paper available at http://link.springer.com/article/10.1007/s10618-014-0358-x. If you want to conduct similar experiments, you can download the datasets below.
Note: 20 Newsgroup is originally a collection of nearly 20,000 newsgroup documents, organized evenly into 20 different newsgroups. It is available at http://qwone.com/~jason/20Newsgroups/.