Preferable Tag Cloud Visualization Formats

Question

Out of curiosity, I would love to know what tag clouds formats best serve the purpose of discovery of more and more (relevant)content?

I am aware of 3 formats, but don't know which one is the best.

1) delicious one - color shading

2) The standard one with font size variations -

3) The one on this site - numbers showing importance/usage.

So which ones do you prefer? and why?

Edit: Thanks to the answers below, I now have much more understanding of tag cloud visualization techniques.

4) Parallel Tag Clouds - a simple use of parallel coordinates technique. I find it more organized and readable.

5) voroni diagram - more useful for identifying tag relationships and making decisions based on them. Doesn't serves our purpose of discovery of relevant content.

6) Mind maps - They are good and can be employed to step by step filter content.

I found some more interesting techniques here - http://www.cs.toronto.edu/~ccollins/research/index.html

Answer 1

I really do think that depends on the content of the information and the audience. What's relevant to one is not relevant to another. If an audience is more specialized, then they will be more likely to think along the same lines, but it would still need to be analyzed and catered to by the content provider.

There are also multiple paths that a person can take to "discover more". Take the tag "DNS" for example. You could drill down to more specific details like "UDP Port 53" and "MX Record", or you could go sideways with terms like "IP address" "Hostname" and "URL". A Voronoi diagram shows clusters, but wouldn't handle the case where general terms could be related to many concepts. Hostname mapping to "DNS", "HTTP", "SSH" etc.

I've noticed that in certain tag clouds there's usually one or two items that are vastly larger than the others. Those sorts of things could be served by a mind map, where one central concept has others radiating out from it.

For the cases of lots of "main topics" where a mind map is inappropriate, there are parallel coordinates but that would be baffling to many net users.

I think that if we found an extremely well organized way of sorting clusters of tags while preserving links between generalities and specificities, that would be somewhat helpful to AI research.

In terms of which I personally prefer, I think the numeric approach is nice because infrequently referenced tags are still presented at a readable font size. I also think SO does it this way because they have vastly more tags to cover than the average size based cloud a la the standard.