How do recommender systems work?
The motivating idea is that we get advice from other people all the time. "Hey,
check out this Web site," "I like this book; you might be interested in
it," "I saw this movie; you’ll like it", or "Don’t go see
that movie!" That happens all the time, and maybe we can formalize the process a
little bit, get some computer support, make it work for us in other situations where we
might not be relying on word of mouth today.
One thing the computer can do is figure out whose recommendations we should pay
attention to, and that’s the first step of the process. Compare which things you
liked and didn’t like to what someone else like and didn’t like, and you get
some kind of score that says you’re pretty much like that person, or you’re not
very much like the other person. Actually, one of the really interesting things is
"you’re exactly the opposite from that person" so whatever they like you
should avoid and whatever they don’t like you should go see. So that’s the first
step. And then, because any one person’s recommendation might not be perfect, you
want to combine advice, aggregate advice, from several people who may be like you. Find
all the people who tend to like the same movies that you like and see which things
they’re recommending and which things they’re not.
Are there some areas where they work better?
The kind of broadbased recommender systems where we just match you up with other people
like you, in general, they’re useful in situations where people have differing tastes
and you want to find someone who has similar tastes to you. And situations where it
doesn’t take a lot of expertise to decide whether you like something. It might not be
the best system for deciding who’s a great doctor, because there are at least some
aspects of being a doctor that you would want medical experts to tell you about. There are
other aspects like, "do they have a good bedside manner?" which is a matter of
taste. Different people like different styles, so you might want to use it for that. In
general it’s better for recognizing things where taste matters.
How do you evaluate how well they work?
You get some suggestions and do you like the ones they suggest... In information
retrieval there are two notions that are often labeled precision and recall. The similar
idea here is, of the things that were suggested to you, did you like them? And of the
things you would have like to have suggested to you, did they all get suggested, or were
some of them missed?
Are some systems better than others?
There are a bunch of different things that use slightly different algorithms. They all
basically have the same flavor I suggested of, first figure out who’s like you and
then combine their recommendations. But the mathematics, the algebra, of exactly how you
decide how someone is correlated with you and how you combine their recommendations,
people have slightly different algorithms. The interesting thing, from my perspective
looking at collection of these that have been tried, is you can find slight differences in
how well one algorithm or another works. But basically if they follow that general
pattern, they work pretty well but not perfectly, and the differences between them are too
slight to matter.
What is the state of the technology
now? Has it fulfilled expectations?
It’s sort of become mainstream, and it’s been used for audio CDs, it’s
been used for books, it’s been used for movies. One place where it hasn’t quite
become institutionalized yet is for newsgroups and bulletin boards, which I think
there’s still possibility for that. That hasn’t happened yet. But if you go to
amazon.com or Borders or Barnes and Noble, they all either have it or it’s about to
be released.
In fact they often have multiple different kinds of recommender systems. If you go to
Amazon they’ve got the version I just described, where it’ll automatically match
you up either based on which books you’ve bought or if you go and say "I liked
this one, I didn’t like that one." But...people can write reviews and post them,
and then they’ve got the official recommended list from Oprah’s book club. A
number of different things, all of which I would call recommender systems of one kind,
although they may not all be as automated as the kind I just described.
Have the companies who develop recommender
system software become commercially successful?
There are a few players in that business. The two that I know about are Firefly and Net
Perceptions. Firefly was bought by Microsoft last summer, although Microsoft made some
curious announcements. Shortly before Microsoft bought them, [Firefly] sold off their Web
site. They had a site that a lot of people were visiting that recommended music. They sold
that off to someone shortly before they sold the company to Microsoft.
Microsoft made some curious announcements about well, we didn’t really buy them
for their recommender software; we bought them because of their work on the Open Profiling
Standard... OPS is a new privacy protocol. And so it appears that they’re not
actually pushing the software that Firefly had; they bought them for another reason.
The company Net Perceptions - I know the founders of it because it came out of a
research project that I was involved in called GroupLens. Their main software product was
called GroupLens, but I haven’t been involved in the company so I have only limited
knowledge of where they stand...
Firefly’s business model was "we’ll run recommender services." At
least that was for a while. They tried some others, but that [model] was when they ran a
Web site. Net Perceptions’ business model was "we’ll sell you the software;
you run the recommender service." Amazon is using the Net Perception software and 30
or more sites have bought the software. They’re running as a "we’ll provide
the plumbing you provide the service."
Who will survive? Will one system eliminate
all the rest?
At least, so far the shakeout hasn’t come that there’s one recommender system
for the whole world. There seems to be recommender systems that are tied in to online
stores and each store has its own recommender system.
The article that Hal Varian and I wrote where we suggested that maybe we’ll end up
with a shakeout and they’ll be just the one - one premise for that is the idea that
there are network externalities. The more people who are using the same recommender
service, the more likely you’re going to find someone who’s like you. Therefore
whoever gets big in the beginning, everyone else is going to want to use their service and
you won’t be able to establish competitors.
I think there are a couple of factors mitigating against that argument. One is that
there are different segments - there’s books and there’s CDs and there’s
clothes and so you could have different recommender services for different kinds of items.
The other is that I think you reach a critical mass. If you have enough people who go to
Barnes and Noble and enough people who go to Amazon, it’s good enough. You can find
somebody who goes to barnesandnoble.com who has the same tastes as you, and if you go to
Amazon you’ll find somebody there who has the same tastes as you. They can both
co-exist; they don’t need to combine, because there’s enough redundancy. There
are enough people who have similar tastes that you could find somebody at either place.
What are the privacy issues?
I don’t think that privacy reasons - people not wanting to have things linked -
were what drove the market to the situation that it’s in today. I think it’s
just that recommender services are an add-on to online shopping. They aren’t the
thing that drives online shopping.
The privacy thing is interesting. One of the interesting pieces about that algorithm I
described for you at the beginning - of we’ll match you up find out who has tastes
similar to yours and recommend things that they liked - that can work even if we
don’t tell you who has similar tastes to you. In some sense, Amazon needs to know
which books you liked, and they can use your recommendations to suggest things to other
people, but they don’t have to tell you who you’re suggestions are going to.
They don’t have to tell the people who are receiving the suggestions who it was
they’re getting them from, so it actually can work anonymously. Sometimes people
don’t want that anonymity and so they have the place where you can post a review. But
you can have a certain level of anonymity. You don’t have to reveal everything in
order for the system to be useful.
Interviewer: But you have to trust the people who are running the
system.
You have to trust the people who are running the system if you want
them to be able to personalize things for you. This is a form a personalization.
We’ll make recommendations of books you’re interested in based on your past
preferences. If you want that kind of personal service you have to say something about
what your past preferences are and you have to trust the company to use them
appropriately.
Is there an industry privacy standard?
Different people have different concerns about privacy. Different sites are better or
worse about protecting it and better or worse even about telling you what their privacy
policies are.
There’s a reason why the folks at Firefly naturally got involved in this work on
standards for information practices. The OPS system that they started working on became
part of the World Wide Web consortium’s P3P - Platform for Privacy Preferences. The
idea is that a Web site would describe what its information practices are - Here’s
the data we collect. Here’s the kind of personalization you’re going to get as a
result. Here’s who we will or won’t sell it to and under what circumstances. We
destroy it after six months. Whatever they do, it’s a way for them to describe their
information practices.
It’s a way for them to describe it in a way that your browser can do something
with it. In particular, your browser can pop up a warning saying "The site
you’re going to has information practices that you told me you don’t like. Do
you still want to go there or not?" ...If you went to a site that promised to use
your data only in the right ways, you might automatically have your browser send your name
and address to that site to make it more convenient for you to enter that data. But you
would only do that at sites that promised to respect your privacy.
That’s all something that’s in the works. Certainly the companies that are
involved in doing these recommendations are taking various positions about what their
privacy practices are, and it’s up to the consumer to pay attention and decide
whether they want to go in or not.
Actually one of the interesting things is that...the other kind of company that’s
interested in this are the online advertising. After you’ve said that you liked
certain movies, the banner ads that you get might end up being for those kinds of movies
or those kinds of books. Another form of personalization is targeted advertising.
There was some concern about people being
scared off by "software agents", as recommender systems are sometimes called,
because the term "agent" had negative connotations for privacy reasons. Is that
still a concern for people?
Things like the term "agent", to describe software that does something on
your behalf, go in and out of vogue. I’m not sure if it’s because of the reason
you just described - because people are scared of things that are called agents. I have my
own feelings about using that terminology... I don’t think it’s a very good term
to use. Not because it scares people because it misleads. It causes people to expect
certain kinds of behavior that the computer program can’t really deliver on, so
that’s where it’s misleading to the consumer.
It’s also misleading to the research or the developer because it focuses your
attention on making a computer program that’s autonomous as possible, rather than
focusing on making a computer program that’s as useful as possible. Often things can
be useful without being completely autonomous. In fact, often they can be more useful.
Instead of trying to guess what people would want the program to do, you make it easy for
people to control what the program does. I prefer to think about user control rather than
programs that guess what users want.
The Internet is good at matching. Why?
The reason why recommender systems have a chance to work better on the Internet than
they do outside of the Internet...is that there’s more people. There’s just more
chance of finding people like you who have opinions about books or movies or bulletin
board messages that you care about. It just let’s us go to a bigger scale.
What’s ahead for recommender systems?
We’ve talked all this time about recommending things by finding people who have
similar tastes. But one of the other interesting features is, maybe those people with
similar tastes are interesting as well. Instead of just using these systems to recommend
things, perhaps we ought to be using them to recommend people.
You could imagine variations of the algorithm that will, instead of matching you with
people who are the same, will match you with people who are complementary. But to a first
approximation people actually are pretty interested in other people who have similar
tastes. At least, if they’re trying to form a book club, you want other people who
like the same books that you do.
I do think that’s one underexploited feature of these systems - not just recommend
things, but put you in touch with the other people who like the same things.
Top of Page
Please direct questions or comments to iota.webmaster@umich.edu.
Last Updated February 5, 1999
