你是否好奇数据可以告诉你什么？你是否想在关于机器学习促进商业的核心方式上有深层次的理解？你是否想能同专家们讨论关于回归，分类，深度学习以及推荐系统的一切？在这门课上，你将会通过一系列实际案例学习来获取实践经历。在这门课结束的时候，

Loading...

来自 University of Washington 的课程

机器学习基础：案例研究

7463 个评分

你是否好奇数据可以告诉你什么？你是否想在关于机器学习促进商业的核心方式上有深层次的理解？你是否想能同专家们讨论关于回归，分类，深度学习以及推荐系统的一切？在这门课上，你将会通过一系列实际案例学习来获取实践经历。在这门课结束的时候，

从本节课中

Recommending Products

Ever wonder how Amazon forms its personalized product recommendations? How Netflix suggests movies to watch? How Pandora selects the next song to stream? How Facebook or LinkedIn finds people you might connect with? Underlying all of these technologies for personalized content is something called collaborative filtering. <p>You will learn how to build such a recommender system using a variety of techniques, and explore their tradeoffs.</p> One method we examine is matrix factorization, which learns features of users and products to form recommendations. In an iPython notebook, you will use these techniques to build a real song recommender system.

- Carlos GuestrinAmazon Professor of Machine Learning

Computer Science and Engineering - Emily FoxAmazon Professor of Machine Learning

Statistics

[MUSIC]

So now, let's talk about how we can use these metrics of precision and

recall to compare our different algorithms that we might think about using.

And to do this we can draw something that's called the precision recall curve.

And let's start by talking about what these curves are gonna represent.

And what the curves are gonna represent is for a given recommender system,

what I'm gonna ask you to do is I'm gonna ask you to vary the threshold on how many

items that recommender system is allowed to recommend to me.

So I'm gonna rank, for example all the baby products on Amazon and

I'm gonna allow you to recommend just one or two or three and so on.

And so that's a threshold that you're varying.

And that's going to trace out this curve.

And what would this curve look like for

optimal recommender where we only recommend products that I like?

Well, what's the precision when I recommend just one product?

Well we know that's a product I like and so my precision,

my world is just that one product, and I liked it, so my precision is one.

And what's my recall though?

If I have let's say, ten items that I liked and

I've only uncovered one, it's one-tenth.

And likewise, as I'm increasing the number of

items I show, my precision always stays at one.

I'm only recommending products I like, but my

recall is increasing because I'm covering more and more of the items that I liked.

So eventually, we will hit this 1, 1 spot.

So the optimal precision recall curve is this line here.

Okay, but let's talk about what the curve might look like for

another more realistic recommender.

I guess before that, let me just annotate that this is our optimal recommender.

Now, let's go on to look at perhaps a more realistic recommender and

what it would look like.

Okay, well, the first product we recommend might not be a product I like,

or it might be, so it's gonna start somewhere on a precision axis.

And then eventually at some point when I vary the threshold enough,

at some point hopefully, I will recommend some product I like.

So both precision and recall are gonna go up.

Then what tends to happen is we add a product that I don't like.

So at that point what happens, well my recall stays exactly

the same because I haven't recovered any more of the items that I'm interested in.

But my precision drops because now I'm looking at a larger world,

a larger set of green squares.

So my precision goes straight down but my recall stays the same,

and you tend to get these very jaggedy looking curves or

you get these drops in precision, then these increases in precision and recall.

And I'm gonna draw it, it won't be completely accurate here, but

it looks kind of like that.

Typically something like this would be a precision-recall curve.

Okay, so this is an example of a more realistic system.

This is another recommender system compared to our optimal.

Okay, so now, that we know how to draw these precision-recall curves, we can talk

about comparing our different algorithms, how do we know which one is best?

Well, we know that we'd like precision and

recall both to be as large as possible and what's the best that it can be?

We talked about the optimal recommender being that curve.

But then when we look at our other curves which are these jaggedy looking things,

one doesn't have to strictly dominate another.

They might do different things at different points.

So one curve is not necessarily, let me actually erase that and

do it in a different color so it's a little bit clearer.

It's not that one is always gonna be better than the other.

Maybe we have another one going here.

So in this case, how do I think about comparing these

different algorithms, and choosing which one is best?

Well like I said, we want precision and recall to be as large as possible but

one thing we can measure to compare these is in general,

which one is doing better than the other and what's a way to think about that?

Well, we can think about the area under the curve.

So we can look at for example, all this area under this blue curve.

And we can compare that to, for

example the area under the green curve, and

I can say which area is larger.

And that is one proxy for

which recommender system is doing a better job.

So that's this point here,

that a metric we can use is something called area under

the curve which measures exactly what I drew below.

But you might not care about how the recommender system is doing

across all possible performance situations.

Instead, you might be in a situation where let's say you have a website and

you know based on the real estate of that page, how many items you can display.

So maybe you can display ten different items to recommend to the user or

you know what the attention span of the users are in general.

And you wanna limit how many products you recommend to 20 products or

something like this.

So in those cases,

where you specifically know how many products you're gonna be recommending,

you care about what your precision is at that number of products recommended.

Because you want that precision to be as large as possible for

the constraint of recommending that number of products.

And so these are two examples of metrics you might use to compare

between different algorithms using this notion of precision and recall.

[MUSIC]