Choice of Attributes and Measures

Loading...

來自 University of Michigan 的課程

Data Science Ethics

40 個評分

University of Michigan

Data Science Ethics

40 個評分

What are the ethical considerations regarding the privacy and control of consumer information and big data, especially in the aftermath of recent large-scale data breaches? This course provides a framework to analyze these concerns as you examine the ethical and privacy implications of collecting and managing big data. Explore the broader impact of the data science field on modern society and the principles of fairness, accountability and transparency as you gain a deeper understanding of the importance of a shared set of ethical values. You will examine the need for voluntary disclosure when leveraging metadata to inform basic algorithms and/or complex artificial intelligence systems while also learning best practices for responsible data management, understanding the significance of the Fair Information Practices Principles Act and the laws concerning the "right to be forgotten." This course will help you answer questions such as who owns data, how do we value privacy, how to receive informed consent and what it means to be fair. Data scientists and anyone beginning to use or expand their use of data will benefit from this course. No particular previous knowledge needed.

從本節課中

Data Validity

Data validity is not a new concern. All too often, we see the inappropriate use of Data Science methods leading to erroneous conclusions. This module points out common errors, in language suited for a student with limited exposure to statistics. We'll focus on the notion of representative sample: opinionated customers, for example, are not necessarily representative of all customers.

Choice of Attributes and Measures6:12

Errors in Data Processing8:22

Errors in Model Design8:01

Managing Change5:18

Case Study: Three Blind Mice4:35

Case Study: Algorithms and Race3:44

Case Study: Algorithms in the Office3:18

Case Study: GermanWings Crash5:50

Case Study: Google Flu5:16

與講師見面

H.V. Jagadish
Bernard A Galler Collegiate Professor
Electrical Engineering and Computer Science

We have to choose what attributes we'll use and how we'll measure them.

And what metrics we will have for our objectives.

And these are important things as we set up our data analysis.

Let's begin by looking at what attributes we choose.

This is something where our choice is usually limited by what is available.

If there are only some properties that we know of the things that

we're trying to deal with, that's what we have.

Now it's possible that additional attributes can sometimes be purchased or

collected, and determining what that might be requires a cost value of tradeoff.

And it's important to think about what attributes are missing?

In other words to make this less abstract,

let's say that we're trying to decide whether to hire someone.

And we look up what we know about them on the web and from their application,

and whatever we know about with respect to their record.

The question is do we know anything about,

let's say what time they go to bed and perhaps,

that's important because of the nature of the job that they're being hired for.

And that is something that maybe one could determine if one had a smart

electric neither in their home from which we could buy data.

But, otherwise,

it's something that is an attribute we probably do not have access to.

There are many choices of this type that a data scientist has

to make in just setting up the problem that they are trying to solve.

There are also attributes that they may decide to leave out.

And sometimes what they're going to leave out is something that's dictated by law.

So for many purposes, you cannot consider the race of a particular individual.

If that is the case, then even if you know a person's race,

that is something that you would not consider

as input to your algorithm or your analysis.

Just as you're making choices with respect to your attributes,

you're also making choices with respect to

the metric against which you are going to evaluate your output.

So let's consider an example.

This is a real-life example.

Kim Kardashian has 15 million followers on Twitter, and

a particular Company X paid her to tweet about its products.

Now here's what happened.

She has 15 million followers,

15 million people presumably saw her tweet about Company X.

1,200 of these visited the company X website.

30 of these 1,200 placed orders worth $30 each on average.

So the company got $900 on sales.

And there's still some question of whether all $900 of these were actually

attributable to Kardashian's tweet.

Maybe some of these would've bought products from the company anyway.

It's also possible there are people who saw her tweet, then immediately visited

the company's website and came back later, and they're not accounted for in this 900.

Let's worry about these things later.

Let's sort of ignore these.

This is a very standard kind of how data analysis gets done.

We just made a big assumption about, yeah, there are these questions.

We don't know the answers to these.

Let's assume those things are zero.

And do I know that's a good assumption?

No, I don't.

And if those assumptions are bad,

my results of the analysis are going to be bad.

But sticking with that assumption for a minute, the question is

if we want to figure out the value of buying Kim Kardashian's time for

this tweet, there are many ways you can do this.

You can say hey, I got 15 million people to have an impression, to see my ad.

Okay, and it's worth let's say a tenth of a cent per view.

So having Kim Kardashian do this for

me is worth $15,000 because she has 15 million followers.

Another way to do it is, it's worth $1 to me for

every new customer who comes through my store.

So Kim Kardashian able to get 1,200 people to visit my website and

since each new visitors worth a $1 that is worth $1,200.

The third way to do it is to say, no,

I want to measure how much profit I got because of this.

So I got $900 in sales and let's say I have a 20% profit margin,

so I got a $180 of extra profit because of this tweet.

And so if I've gotta do a contract and the question is,

how much I will be willing to pay for her to tweet about my company's product.

The answer could be anywhere from some fraction of $180 to

some fraction of $15,000.

Depending on what the measure is that I am using as the thing that I am looking for.

Well, here I kind of set up this problem to make clear that what

one was really looking at is the sales and often one may not have that and

therefore one is pricing things on things like page views or

impressions and the results can be astonishingly different.

探索我們的目錄

免費加入並獲得個性化推薦、更新和優惠。

Coursera 致力於普及全世界最好的教育，它與全球一流大學和機構合作提供在線課程。

© 2018 Coursera Inc. 保留所有權利。

通過 App Store 下載

通過 Google Play 獲取