I am extremely impressed at authors' ability to cram so much information into such a short paper. However, I think due to it's density, there were quite a few details that were assumed to be known or simply left out. So, here are a list of questions/comments that I had:

In section 2.2, what is a reference partition?

Why did the authors decide to use k=3 for the k-nn classifier?

For the unconditional mixture modeling, do you simply threshold the mixture overlap probability? If so, how is this threshold chosen? The authors seem to ignor this in the description and the experimental section.

I was a bit disappointed with the experimental section. I found it a bit difficult at times to understand exactly what was happening. It doesn't appear to me that the authors used cross-validation in their experiments. If they didn't, why didn't they? In the first set of experiments seen in Figure 2, are the feature selection methods applied sequentially or are those results of the algorithms run independently? The experiment in section 5.2 is a bit confusing for me. Although I think it is interesting to show how the classifiers perform with a consistent feature set, I think it also would have been interesting to see how the classifiers interact with the classifier dependent feature selection algorithms.

Finally, given that one of the motivations for applying filtering methods initially to narrow the feature selection problem is to minimize computation time, I think the paper should have presented some notion of computational complexity improvement even if it was only in the form of experimental observation.

Sorry for the long post... I had a few questions :)