Intelligent agent sites

On 13-Sep-06, at 14:34, maclistsadditional.com wrote:

>
> And so, I'm wondering: has anyone found one of these sites that works
> well? It seems like such a promising idea.
>
> Dave
>
> [1] Most interesting so far: <http://movielens.umn.edu>
>

I agree it is promising, so far the one I liked best is http://
www.pandora.com/ but as I do not live in the USofA, I was unable to
use it for long. What was nice was that my feedback was immediate -
only three inputs, but good ones: I don't like this, I like this,
it's ok but not for a month. But, it is a music site so input can be
immediate. I would hate to spend money on something based on the how
I felt about the songs I heard. If you 'joined' it looked like it
would be better in a fairly short time.

Mary Arthur

"Dave" wrote to TidBITS Talk:
> Aside from the humourous result from assuming that linking authors is
> enough to conclude that someone who bought Adam's wireless networking
> book would be interested in a book on SPSS, it got me thinking about
> the various attempts to find common ground. I've tried a couple
> sites [1] that recommend news articles, movies & music based on my
> declared preferences, but the results were, at best, well, mediocre.
> And so, I'm wondering: has anyone found one of these sites that works
> well? It seems like such a promising idea.
> [1] Most interesting so far: <http://movielens.umn.edu>


A pet peeve of mine, and part of my job! Caution, long post ahead.

At the CNI fall 05 task force conference, I attended a briefing on the Techlens project that you link to, and there were some interesting observations on when recommenders work, and when they don't (most in the Q&A, so not covered in the abstract).

http://www.cni.org/tfms/2005b.fall/abstracts/PB-techlens-konstan.html

Primary problem of a recommender system is the problem of neighbourhood. How widely do you, as user, want to have the recommendations vary? When you are new to a subject, you want the defining standard works - a narrow view. As you get more versed in the subject, you actually don't want those predictable results anymore, as you will be already familiar with them. Without surprise, it has no value. Different users want different results.

Research on users expectations showed that users were most content with a recommender service if it would give 5 suggestions (in an unobtrusive interface), as long as out of these five one or two would be 'interesting'. Keep in mind though that this was research on users in a strictly defined research field, which can't be translated directly to other fields, but it gives an indication.


How does this translate to amazon? Like Dave, I get the occasional amazon suggestion by email, most of which I delete instantly. Only rarely they were actually interesting. As a result, I find them annoying or amusing, depending on the actual suggestion ;-) However, when I browse amazon, the recommendations are much less obtrusive, so I glance at them when I want, and then I sometimes do find something interesting in there. And I find myself agreeing with the outcome of the techlens research: my amazon miss:hit ratio is 25:1, and I would like more hits, but it needn't be 1:1.


As for the suggestions themselves, they depend on the quality of the data. The ACM techlens used citations to see which objects were linked; Amazon has to rely on more primitive metadata, such as the author. Because it refines this with buying/browsing patterns, it is IMHO actually pretty good, but as with all 'social software' (of which amazon is the granddaddy) this needs a critical mass to get reliable.

Returning to the original question. A good recommender system will always give you some surprising suggestions. It may not always be the surprise you wanted, but if it would be predictable, it would be of no value at all! So by definition, there is a high miss to hit-ratio; therefore, a system must be unobtrusive to not appear, in the words of the original poster, mediocre.


In the long run, this will all change, when the systems will be able to parse the objects (book, movie, etc.) and build relations on the actual content. There is a lot of research in this area, largely spin-off of 'Homeland Security' projects. But it is still years away.


--

____Driek_Heesakkers________________________________________________
Senior Projectleader, University Library, Universiteit van Amsterdam
 E: driek.heesakkersuva.nl T: +31 20 525 4966 W: www.uba.uva.nl

  Lisa: My family never talks about library standards. And every
        time I try to steer the conversation that way, they make
        me feel like a nerd. [http://www.snpp.com/episodes/AABF18]

On 13 Sep 2006, at 14:34 , maclistsadditional.com wrote:
> I've tried a couple sites [1] that recommend news articles, movies
> & music based on my declared preferences, but the results were, at
> best, well, mediocre.

I like pandora.com quite a lot.

--
"Here comes sunrise. Yeah, here's your sunrise. I used to hide from
the sun, tried to live my whole life underground, why'd you have to
rise and ruin all my fun? Just turn over; close the curtains on the
day."

Pandora's system is different; they analyze music and recommend similar music. Amazon's system has more to do with buying patterns, what you bought and/or like, what others bought and/or like, where these things overlap, etc. in addition to related items.

I feel Amazon's system's quite good, though I don't let them e-mail me recommendations. ;-) I used the "___'s Store" link to browse recommendations for fiction & DVDs. I believe it can be useful if you manage it, and if you remember a computer isn't a substitute for a personal recommendation.

My top recommendation for managing Amazon's recommendations: If you buy something unusual [for you to buy on Amazon], something as a gift, or something you don't want recommendations related to...then mark it to not be used for recommendations. This simple checkbox is very useful. Go into "___'s Store" & click on "Items you own."

So I uncheck "Use to make recommendations" for anything other than DVDs and fiction (books). (I find music tastes vary too wildly.)

Two other things you can do to improve recommendations (or so Amazon sez):

1. Rate items, especially ones you own.

2. Check "I own it" by items you own (in categories where you want recommendations), unless it's a very atypical purchase. For example, for books, I read mostly F&SF; I don't check the occasional mystery or YA book I own or buy, because the recommendations aren't generally right for me.

You can mark items showing up in the recommendations as "not interested," but I'm not sure their system actually learns anything from that. I believe it just means Amazon won't recommend that item to you; it would be nice if it learned from this, I suppose. (Similarly, I'm not sure if rating something 1 star means they recommend fewer items like that...or if rating things only has a positive affect, e.g., perhaps only rating something 4-5 stars makes a difference by including more items like that item.)

I've also noticed that occasionally things go wacky and my recommendation list will change a lot from one day to the next. My wild guess is that Amazon periodically adds new information, cross-referencing purchases, etc. -- not continually.