::scr googley eyed

[Ben]

On Tue, Oct 15, 2002 at 10:53:24AM +0100, Simon Wistow wrote:
> 
> For a freelance project I've been looking at collaborative filtering -
> the technique of learning from a community to provide better
> information. PageRank is an attempt at this. A ver simplistic one but an
> attempt. However I think there are better ways to do collaborative
> filtering. 

As I may have mentioned / ranted about, collaborative filtering only works in 
places where you have enough of a repeating 'market'. It seems to me that
certain types of searching are indeed this, but it's far from clear that 
*all* types of searching are.

Berrypicking is a repeating market because of the iteration involved in
refining the query. Whether CF can give a good answer to the question of *how*
queries evolve towards resolution is interesting.

Known-item searching is interesting too. In the case of a finite information universe,
like Amazon, CF should work because the questions that a user *could* ask
are fundamentally limited. Amazon's recommendation engine is worth a consider
or two, by the way. It's pretty accurate, but not very adventurous, and when 
I introduce it to one of my interests which is outside its experience, it tends to panic
a bit, and have to be retrained.

After seeing nothing but industrial and goth records, it got very confused when
I added a Gilbert and Sullivan record to my wishlist, and suddenly started
suggesting me mouldy cast recordings of 'Cats'. No thanks. I eventually had to remove
the G&S and calm it down with Front 242 for an age.

Ben