I agree with all the criticisms below, but like everyone else, I think the paper presents a very good idea.

The opening statements were the most interesting to me. Several of the papers have addressed this obliquely, but this one is very direct. Why approach the problem sideways?

I think the paper would be more useful with more time spent on the algorithmic details and explainations. It is not sufficient to use many unexplained terms and expect the readers to automatically understand them. The heavy use of acronyms also was difficult to parse. I did like the extensive experimentation with many standard datasets. The broad range of tests ensures that members of the machine learning community have worked with at least one of those test sets, and can compare their results.

Melanie

I agree with Sameer in that some of the details of the paper, including the algorithm, seemed a bit sketchy. I thought the paper did a fairly good job, however, of presenting experimental results. Most of the time I felt I knew fairly well what the experimental set up was and I liked that they gave error bars, particularly since the differences between algorithms was so small. I had a couple of minor complaints with the figures, though. In figure 2c, they change which axis NB+ILQ is on, which was confusing at first. In Table 1, the paper presents a large amount of data. It would be helpful if the paper highlighted or underlined important or significant data.

Dave

hi,
I really like this paper for its idea.
and dislike it very much for its use of terms without explaining what they mean.

Considering the fact that NB classifiers are effective precisely because even though their probability estimates might be wrong their ranking are correct in many cases (since its easier to do discrimination than density estimation), maximizing conditional likelihood makes a lot of sense. I think this is an idea well worth exploring in other the context of learning algorithms also.

The results look impressive but the details on the algorithm are sketchy. I guess conference page limits do that to papers.


sameer