W3C home > Mailing lists > Public > ietf-http-wg-old@w3.org > September to December 1997

Re: Quality factors

From: Martin J. Dürst <mduerst@ifi.unizh.ch>
Date: Thu, 9 Oct 1997 14:52:15 +0100 (MET)
To: Graham Klyne <GK@acm.org>
Cc: cuckoo.hpl.hp.com@http-wg.uucp
Message-Id: <Pine.SUN.3.96.971009143357.7026q-100000@enoshima>
On Thu, 9 Oct 1997, Graham Klyne wrote:

> I suggest that in this case a 3-digit (max) number is insufficient, as with
> a significant number of alternatives an implementation will soon run out of
> space within which to slot further entries between existing entries.  I
> estimate that a perverse presentation would run out ranking space after
> about 10 entries (log2(1001)).

Hello Graham,

Your arguments give one important aspect of the problem. There are
others.

The fact that only a coarse granularity should be used in
case of Laguage preferences to avoid giving too much traceable
information has been discussed here already.

Various aspects of quality are usually combined by multiplication;
this means that each q value has to be appropriately scaled
(e.g. if for languages, fr has q=0.6 and en has q=0.8, and for types,
we have q=0.5 for text/html and q=0.7 for text/pdf, then a French
PDF wins against an English HTML). Note that this scaling
also allows to change the importance of each factor, by using
exponents (if we want language to be twice as important as type,
we square the q's for language, get q=0.36 for fr, and q=0.64
for en, and now the English HTML wins against the French PDF).

From a user perspective, I would like to claim that in the higher
quality area, fine distinctions are more important and easier than
in the low quality area. They are more important because there is
a higher probability that one of these will actually be choosen,
and easier because the quality differences are actually relevant
to the user in that area. This suggests that in your case, where
you have to assign quality sequentially, you could divide intervals
asymetrically (e.g. 70%/30%) instead of half-half. So the first
document would get quality 30%, the second would get 9% if it is
lower and 51% if it is better, and so on. This would allow you
to rank differently almost 20 documents if each subsequent one
turns out to be even better than the previous one. If the
subsequent ones turn out to be worse, you won't be able to
make any distinctions anymore after the 5th or so document,
but that won't be too bad.


Regards,	Martin.
Received on Thursday, 9 October 1997 07:59:05 EDT

This archive was generated by hypermail pre-2.1.9 : Wednesday, 24 September 2003 06:33:01 EDT