[lug] Weighting of factors and whatnot (a bit OT)

[Steve Sullivan]

Chris,

Two other approaches you might look at ...

Cluster analysis in 6D space.

SVM == support vector machines, and their more recent cousins,
minimax theorems.  These are useful for forming binary
separations - good versus bad, for example.  But the separation
is created by a single numerical value that could be used
for your single number.

The classic paper is:

A Robust Minimax Approach to Classification
Gert Lanckriet, Laurent El Ghaoui, Chiranjib Bhattacharyya, Michael I. Jordan
Report UCB/CSD-02-1218,
EECS, UC Berkeley


Steve

> From: "Chris Riddoch" <riddochc at gmail.com>
> Subject: [lug] Weighting of factors and whatnot (a bit OT)
> 
> So, this isn't entirely a linux thing, but I figure there are enough
> programmers here that I'm likely to get some idea of a good solution
> to an interesting problem.  The actual thing I'm trying to solve has
> nothing to do with books or genres at all, but this is the closest
> metaphor I can think of to the real problem as I've been able to come
> up with.
> 
> I've got a list of things (call them books) I want to rank.  Each of
> the things to rank has six numbers associated with it, one for each
> category (genre).  Each of these numbers represents the degree of
> membership of the book in a particular category, from 0 (not in the
> genre at all) to 1 (unquestionably belonging in the genre).
> 
> Now, the "ideal" book belongs mostly to a particular genre, slightly
> less so to another particular one, less so to the next, and so on to
> being not at all part of the least desirable genre.
> 
> I need to distill the factors into a single number that can describe
> how close to "ideal" a book is, given a desired ranking of genres.
> Note that I don't have any specific desired degree of membership for
> each genre - I don't mind any two of them are especially close to each
> other, so long as it's not (for example) the most and least desirable
> genres.  That final, single number, will be the basis of the ranking.
> 
> It seems like a class membership, set theory, perhaps permutation sort
> of thing.  I don't have any books on this topic.  I suspect there's
> more than one way to do this.  Does this problem remind you of any
> similar known problems with known solutions?  What do you suggest?
> 
> -- 
> epistemological humility
>   Chris Riddoch
> 
-- 

========================================
Steve Sullivan    sullivan at mathcom.com

http://www.mathcom.com    303-494-7115
========================================