I thought that the idea of including information from linked sites very interesting.

I particularly liked the introduction of the paper. They give a good chain of ideas that lead to the development of their system as well as the motivations for each step that they take. The ideas are explained in a general way so that even with minimal knowledge, a good idea of the paper’s goals and methodologies can be obtained.

Another nice thing that the paper does is build on existing research. The paper builds up their system from a general text classifier. The nice thing about this is that any improvements to basic text classifiers will result in gains in their system with little or no work.

I found some of the experiments a bit confusing. Their first experiment tests three different cases. In the third case they distinguish between local and non-local data. I am not exactly clear how they actually used this information, though.

I liked how the paper showed the performance increase with an increase in knowledge about the neighbors. I thought it was quite interesting that they go such good results from the prefix+Link even without any neighbor classifications. However, I would have appreciated some discussion of the likeliness or difficulty of actually having that many neighbors classified correctly. When we are dealing with such large datasets, even pre-classifying some small percentage may be extremely time consuming.

Dave

I agree with David about the interesting ideas introduced in the paper. I liked the idea of text classification in a global context, which exploits the natural taxonomy of the web.

I felt that the ideas and concepts were clearly explained, but implementation details were neglected, and running time measurements or analysis was missing form the paper.

The writing style of the paper was clear, however in certain places the intoduction of new variables and notations was a bit confusing and hard to remember.

Overall, i found tha paper very interesting, but would like to see some performance results in terms of running time.

The idea of using links as features are very interesting. In that way, they can use their previous work as much as possible
I think the formulas in the paper is not that clear. I believe they are not very related to this paper.
The graphs looks ugly though.

I simply don't understand some formula.