neural networks and filePro

Bill Vermillion fp at wjv.com
Tue Nov 16 18:00:28 PST 2004


On or about Tue, Nov 16 20:31 , while attempting a Zarathustra 
emulation Fairlight thus spake: 

> This public service announcement was brought to you by Bill Vermillion:

...


> > And there is a section called "Improve my recommendations" base on
> > previous purchases so you can sort of fine tune your
> > recommendations.

> I'm already using three of their RSS feeds. I don't need to
> clutter my inbox as well. :)

You don't with 'improve your recommendations'.  You rate you
preference for items you have, or are in you wish-list, or hold
box, with a score of 1 to 5.

> > You can change that too. And it's not on personality but
> > on previous purchases - and you can exclude items from the
> > recommendations.

> Yes, but I've seen some dreadfully mismatched recommendations
> based on past purchases. :)

And mine have been quite good.   If you want them to be better you
have to input a little something - otherwise all your purchases
will be ranked evenly.  And if you bought something for gift
you don't want that ranked.  When you click on 'why was I
recommended this' you will find your previous choices/hold items
used, and you can check the ones off that you don't want to
be taken into account.  If you don't give any feedback you get what
they give you - so don't complain.  They can't read your mind.

> > > Hell, people can't even accurately judge others'
> > > personaOlities--I surely wouldn't trust such a subjective
> > > job to a computer, which will be easily deceived, and which
> > > has no concept of motivations.

> > Why do you think it is judging personalities. It's based on
> > items you have purchased.

> I don't necessarily say it -is-, I'm thinking in terms of
> Walter's original post. When I think neural network, I think of
> the capacity for learning.

prrect/

> I don't necessarily say it -is-, I'm thinking in terms of
> Walter's original post. When I think neural network, I think
> of the capacity for learning. I don't think it should be raw
> correlation of, "You bought 'x', 20 other people bought 'x'
> and also bought 'y', so you would probably like 'y' as well."
> That's a broad generalisation based on little data.

But from my observation based on the data point of one person - me
- it appears things are deeper than that.  It's not "joe liked this
and you liked it too - so if joe likes that you will like it".

It appears there are deeper categories - and I can't tell you what
they are - but sometimes I'll get a recommendation on something I'd
not heard of and look at it - and son-of-a-gun - it looks
interesting.

As an example I was recommended "Rules of the Game".  When I
checked to see why it was recommend it was because I had
purchased Kieslowski's "Decalogue", "Un Flic" with Alain Delon,
and "The Damned" with Dirk Bogarde. And I had also marked
"Andre Rubelev" as one I might want to get.

And in an email I had mentioned "21 Grams".  I see that is on a
recommended list because of "The House Of Sand And Fog", the
bizarre film "Masked and Anonymous" with Bob Dylan, four Tarantino
films, and a couple of others.

You do NOT have to buy selections but you can rate things totally
independant of purchases or want lists and those will be used
to select items.  So the more you put into it the better it gets.

I'm constantly surprised at how well their system works - but then
again I've rated things I've seen elsewhere to help judge my
tastes.

> A -real- neural net I should think would have at least a
> limited ability to "learn"...at least some form of AI would
> be involved. So it would look at -all- the purchases someone
> has made and develop personality profiles and -then- compare
> personality types rather than raw purchases.

Personality types should have nothing to do with what a person likes.
Of course that depends on what you consider a personality.

> It's sort of
> like the MMPI-II, but applied to "normal" (whatever that is!) individuals.
> You take a standardised set of criterion and apply it over a -broad- range
> of questions that cover critical areas, and you can evaluate someone's
> personality fairly accurately based on the results.  But you can't do it on
> the basis of two people answering questions 23 and 37 identically and one
> answers question 37 with a 'C', then the other likely will have as well.
> Or worse, generalising a -lot- further by genre--you buy one type of hard
> SciFi and get recommendations for a bunch of much weaker SF, or a really
> specialised side of Fantasy (Glen Cook's Black Company series) that is
> -not- your classic Fantasy, and you'll start getting recommendations for
> all sorts of regular old Fantasy titles.

> It's just not intelligent enough.

You're opinion - not mine.  And I suspect it's because you haven't
ever bothered to make ratings so it could fine tune them to your
needs/wants.

The recommendations on music and books can be strange - because
my purchases have been strange.  Then I also get occaionaly emails
from amazon.co.uk - as I've bought things there that aren't
available here.  Same with amazon.ca.   The 'net does away with
borders - but the exchange rate for English purchases is getting
to be a bit much, as the dollar shrinks futher in relationship to
the pound.  It is nice that the one registration will work at
any Amazon site anywhere in the world.

And that's the last I'm contributing to this thread.

Bill

-- 
Bill Vermillion - bv @ wjv . com


More information about the Filepro-list mailing list