SIGIA-L Mail Archives: Re: SIGIA-L: Popularity in Relevance Ran
Re: SIGIA-L: Popularity in Relevance Ranking Was: understanding metadata
From: Peter Merholz (peterme_at_peterme.com)
Date: Fri Sep 07 2001 - 11:50:38 EDT
> So, the
> popularity approach can sometimes give false legitimacy to mediocre
> information.
Sure. This positive feedback loop can occasionally lead to the heightening
of mediocrity.
This has been a common criticism of such self-organizing and
collaborative-filtering systems...
But, I think it's important to remember that, *by-and-large*, popularity is
an extremely useful attribute for most of the world most of the time. We IAs
are an odd sort, searching high and low for special nuggets of information
to scratch particularly narrow itches. For most information retrieval, this
simply isn't the case.
Still and all, there are ways of mitigating this factor. Two occur:
a) Randomness. Any good self-organizing system needs to incorporate an
element of randomness in order to ensure the evolution of that system. It's
random mutation that spurs biological evolution... And most of the time, the
mutation is a dud and not selected for. But occasionally it's a winner, and
the system becomes more fit.
b) Personalization. This was what Epinions is trying. Imagine if Google knew
more about *you* and the kinds of sites you visited, and what you thought of
the quality of those sites. Then, when you searched for "burger," it would
take you to a burger site that better suited your tastes (probably through
some collaborative filtering means). If someone else entered "burger",
they'd be offered a burger that better suited their tastes.
The problem with such personalization is that it takes a long long time for
a system to get enough information about 1) you and 2) others like you in
order to offer quality results. But Amazon, I think, is a case in point of
this working quite well... After collecting purchasing and rating data on me
and millions of others for the last 4 or 5 years, it offers me some
scary-spooky apt recommendations.
--peter
This archive was generated by hypermail 2.1.2
: Sun Nov 23 2003 - 22:54:48 EST
|
