This course, which is designed to serve as the first course in the Recommender Systems specialization, introduces the concept of recommender systems, reviews several examples in detail, and leads you through non-personalized recommendation using summary statistics and product associations, basic stereotype-based or demographic recommendations, and content-based filtering recommendations. After completing this course, you will be able to compute a variety of recommendations from datasets using basic spreadsheet tools, and if you complete the honors track you will also have programmed these recommendations using the open source LensKit recommender toolkit. In addition to detailed lectures and interactive exercises, this course features interviews with several leaders in research and practice on advanced topics and current directions in recommender systems.
Predictions and Recommendations
Loading...
Из курса от партнера University of Minnesota
Introduction to Recommender Systems: Non-Personalized and Content-Based
303 оценки
Из урока
Introducing Recommender Systems
This module introduces recommender systems in more depth. It includes a detailed taxonomy of the types of recommender systems, and also includes tours of two systems heavily dependent on recommender technology: MovieLens and Amazon.com. There is an introductory assessment in the final lesson to ensure that you understand the core concepts behind recommendations before we start learning how to compute them.
Познакомьтесь с преподавателями
Joseph A KonstanDistinguished McKnight Professor and Distinguished University Teaching Professor
Computer Science and Engineering
Michael D. EkstrandAssistant Professor
Dept. of Computer Science, Boise State University
Welcome back.
In this lecture, we're going to talk about predictions and recommendations or
more specifically about the output of a recommender system.
So our learning objectives are to understand different ways on which
recommenders produce output and that output can be used.
To look specifically at the distinction between prediction and recommendation.
And the distinction between explicit and
organic presentation of those predictions and recommendations.
And we're going to do all of this by reviewing examples that may help us
understand which presentation makes sense in a different situation or
different application.
So, the idea of a prediction is a simple one.
In many recommender systems,
what's being computed is an estimate of how much you will like an item?
Or how likely you're to buy an item or
some other numeric score that's supposed to correlate with your opinion of an item.
These are frequently scaled to match your rating scale, so
if you're rating from one to five, you may get back one to five stars.
And they're often tied to search or browsing for specific products.
So to give you some examples of this In the Zagat guide,
we looked at a couple of lectures ago,
the place this shows up is in the numeric scores.
If you come over here, you're getting the food,
decor, service, and cost predictions.
These are not on the same scale.
When it says decor twenty-seven, it's really saying,
we predict you're going to be closer to three than you are to two.
Even if you look at Coursera,
you'll see in this area a set of indicators
that are accumulating measures of how much
the system thinks you might like this course.
In a very non-personalized generic way,
3.6 thousand people on Facebook liked this course or
202 people plus oned it on Google Plus.
Or 536 people have tweeted this course, that's an indication of
how much you might like it and again it's a form of prediction.
In this case, not reduced at all to a scale In MovieLens,
which you'll remember from our introduction, the prediction is given
in the star scale right over here.
So The Burbs is giving a prediction of three stars,
whereas Apollo 13 is getting four and a half stars.
What you may notice is this was actually linked to a particular search.
The search I did at this point was find me movies starring Tom Hanks.
And then, I asked it to order those movies alphabetically and
I could then browse through them and see which movies I might want to see,
if I were looking to see a Tom Hanks movie.
And if I believe these predictions,
then I probably should go see Apollo 13 or Band of Brothers,
rather than seeing The Burbs or Bonfire of the Vanities.
Recommendations don't make the bold statement that predictions make.
With the prediction as you saw, came back and
said look on a five point scale Band of Brothers is a four and a half.
There is a risk, that if I look at that and
say, you've got to be kidding I hated that movie.
I may distrust this recommender because it made that prediction in an angered and
in a way that I now lose trust.
Much as we talked about people feeling like, if Zagat gives a 23 to their
favorite restaurant, they don't believe the guide anymore.
With a recommendation, we're making a suggestion of things you might like.
Things that fit.
What you're doing right now.
These are often presented in a form of a top-n list.
Top ten, top five, sometimes they're just placed in from of you.
So if you search Google Images for me,
these are the things it thinks I might be interested in.
I'm delighted that the first one is me.
The next two are my uncle.
There are a few other Konstans that I'm not related to.
There's some there maybe I'd like to be related to, I don't know.
But It didn't come out here and say, this has a 93%
chance of being what you're looking for, and this one has a 90% chance.
It just gave me the items.
If I go back to MovieLens and say, look,
what I'd really like are science fiction movies, and
don't show me the predictions because I don't want to be swayed by them.
Just show me science fiction movies from 1974.
And I also selected let's do those in English.
So that you can see I've said, Sci-Fi,
in English 1970 to 2013, and
don't show predictions.
Then it will give me back, effectively, a recommendation set,
because we think you'd like Minority Report, Spider-Man,
and Inception, and V for Vendetta.
And there's no prediction to be seen, now this is purely a recommendation.
Or there's lots of other places if you go to the electronics store Newegg,
you can see top selling smartphones or
the top unlocked budget phones, or the newest unlocked phones.
Again, recommendation lists tied to a particular interest, or activity.
Doesn't say, we think you're going to like this.
Just says, hey, if you're interested in unlocked budget phone, these are top ones.
And in this case, it doesn't even say what top means.
Top reviewed, top selling.
We make the top profit something like that.
We don't have to know because we're browsing and
we might choose to explore them.
Now predictions and recommendations often do come together.
What we need to understand is that there are pros and cons to each.
The nice thing about a prediction is it helps quantify,
it gives you a very clear understanding on some scale.
How good is this?
How much do you think I will like it?
In the Zagat world, if that restaurant says,
28, 39, this is saying, hey,
this is worth a special trip, it says 25, 26, 27.
This is a great restaurant, you won't necessarily travel for it.
It says 13 might be a utilitarian place to go to but nothing special.
In movie land it becomes back and says four and a half,
five stars this is a rare things that it says you're going to like a lot.
It comes back with three and a half, well, you probably won't dislike it.
You might like it but it's not special.
The con with predictions, the real downside is that it gives you
something that can be wrong to use the scientific term they're falsifiable.
As soon as somebody sees hey you said this was four and
half stars and then they look at it and say but its not.
You're wrong.
And in many cases, particularly when we're talking about selling people things or
trying to getting them to consume information.
It's better to not give them a chance to say your wrong,
that's why many search engines just give you a list.
Many of the retail sides also just put out a list of items.
Recommendations the pro is it gives you a set of
goof choices It's a set of things to give you.
The con is, if it's perceived as a top-n, if it's perceived
as a weighed list of recommendations ranked from best to worst.
As soon as someone starts seeing ones that are bad they may stop exploring.
In a search engines, this is known as the if it doesn't fall on the first page,
it's lost phenomenon.
People look a few down, and then when they start to get less relevant, they're done.
And this can happen in lots of different domains, so
if you'll take the example here of looking at an ice maker.
This is an example that we have from
Amazon's email to me of a recommended set of products.
As I look at this thing and I say, well, that first one looks sort of potentially
interesting and the second one looks okay.
I may not go very far down the list before I stop and
that could be a reason that I don't want to go much farther
because this has been presented as some interesting items.
But I may have interpreted as these are the best fits for you.
Another example, that you'll see where things are not really ordered in any
particular way is Zappos.
Where it's just simply coming out and saying, here are some products.
Here are some shoes, take a look.
These are shoes that happen to fit a size But they're not best, they're not worst.
And that leads us to the last dimension we're going to consider today.
How explicit is the recommendation?
We use the terms explicit versus organic this is not a term of art in the field,
this is attempting to be descriptive.
And this comes back from back a long time ago.
You go back to the mid 90s when recommender systems were just
becoming popular commercially.
Companies spent a lot of money to buy recommender system to
personalize their commerce offerings and their content offerings.
In our work with Net Perceptions, we were working with lots of companies that
were spending tens of thousands of dollars and more.
To implement a recommendation that would pick a product for
you whether that was a piece of art, music, books or
anything else that they might be trying to sell.
After making that investment,
many of these companies didn't want their costumers to fail to notice.
That they had improved their site this were competitive times and
so the sites would come back and say, here something we picked just for you.
That's an explicit recommendation sometimes they'd even
put a score out there and say, we picked this just for you.
It's 4.5 on the scale of 5, that's an explicit prediction.
Now, explicit has its benefits,
it did call attention to the recommendation or prediction.
It also had its drawbacks.
Customers often felt they were being pushed towards something.
They were being manipulated.
They didn't always trust the recommender and a softer presentation
might have been a little bit better some of you have had this experience.
If you go into a store and a salesman or
saleswoman gets right in your face saying, well I have the right one for you.
Before learning very much about you.
No, this is the right television, I know you want our new 60 inch LED.
You have a natural reaction to push back and say that's not what I wanted.
You don't know me that well.
All I said was, I'm looking for a TV, you must be trying to push that for a reason.
On the other hand, an organic presentation putting
the TVs out on the shelf, and putting the ones that are ones who
think are likely to be attractive on the end cap, or on a prominent position.
That might use the very same recommender, but its a softer sell.
In book stores, you see this when some books are facing forward so,
you see the whole cover of the book.
And other books you only see the spine,
that's not accidental that somebody making a recommendation.
Somebody saying, hey, this is a book I think you're going to be looking for,
so we're going to make it easy to find, or
this is a book I think you might be interested in buying.
So we're going to put it here where it can remind you that it's available for
purchase.
So when you look today, you'll see that sites that are out there strike a balance.
There is definitely explicit prediction going on in many sites but
it's pitched not so much as we think this is what you'll think.
But we think you'll find it useful to know that other people gave this four stars.
We think you'd find it useful to know that this is the best seller in the category or
number three in the category and
we're seeing that balanced with some coarser granularity recommendations.
All over the page, as we saw with our Amazon tour,
are things that are just placed there with little more than you might like this.
As a way of trying to put that recommendation in a more subtle organic
presentation.
And that balance between these are the best and
the softer presentation, here are some things that'll
be interesting is the balance that we try to strike.
So you should now understand the difference between prediction and
recommendation.
The difference between showing somebody a predicted score and
giving somebody an item or
a set of items that might be useful without trying to predict that score.
You should also understand the range of presentations that are out there
that you can give a prediction that's very hard and explicit.
You'll think this is four and a half stars.
You can give a prediction that's been soften a bit
to say hey other people think this is four and a half stars.
Just so you'll know, you can have recommendations that say hey,
this is the top five, and
you can have recommendations that just put five products on a page.
And that balance between organic and
explicit is one that you have to strike as you design an application.
And you should understand that there are advantages at every point along that
spectrum.
There isn't a single correct that has to fit your application.
Ознакомьтесь с нашим каталогом
Присоединяйтесь бесплатно и получайте персонализированные рекомендации, обновления и предложения.
Coursera делает лучшее в мире образование доступным каждому, предлагая онлайн-курсы от ведущих университетов и организаций.
© Coursera Inc., 2018 Все права защищены.
