#274: Real Talk About Synthetic Data with Winston Li

欢迎来到数据分析强力时段。

Welcome to the analytics power hour.

我们会以轻松对话的方式聊数据分析，有时也会用到直白的表达。

Analytics topics covered conversationally and sometimes with explicit language.

大家好。

Hey, everyone.

欢迎各位。

Welcome.

现在是数据分析强力时段，这是第274期。

It's the analytics power hour, and this is episode 274.

但今天，我们要深入探讨一个话题。

Yet today, we're diving into a topic.

听起来像是科幻剧本里的情节，但实际上它完全根植于现实中的数据世界。

It sounds like maybe it came from a sci fi script, but it's actually very much part of the real world data landscape.

没错。

That's right.

合成数据。

Synthetic data.

不管你是用它来做机器学习、保护隐私，还是只是为了让仪表盘在真实数据不配合的时候有点东西可处理，它现在在我们这个行业里可真是火了一把。

You know, whether you're using it for machine learning, protecting privacy, or just giving your dashboard something to chew on when the real data won't play nice, it's definitely having a moment in our industry.

和原始数据不一样，它不会突然失踪、丢值，也不会冒出一堆莫名其妙的异常值，更不会出现那种像有人睡着了乱敲键盘一样的数据行。

And unlike original data, it won't ghost you with missing values or weird outliers or those inexplicable rows that look like someone fell asleep on the keyboard.

我们会聊聊它是怎么生成的，在哪些地方表现突出，又在哪些地方可能不太靠谱。

We'll talk about how it's made, where it shines, and where it might fall short.

所以不管你是不是数据科学圈里的老手，或者只是对数据感兴趣，我觉得这期节目你都会喜欢。

So whether you're deep in data science or just data curious, I think this podcast will be for you.

这期节目就像合成数据一样，希望能有明确的目的，而且出乎意料地实用。

It'll be kinda like Synthetic Data, hopefully generated with purpose and surprisingly useful.

不过在那之前，先让我介绍一下我的搭档，瓦尔·克罗尔。

But first, let me introduce my cohosts, Val Kroll.

最近怎么样？

How you going?

你最近怎么样？

Or how you doing?

我太习惯介绍瓦尔了。

I'm so used to introducing Val.

不是莫。

Not Mo.

我知道。

I know.

你最近好吗？

How are you?

我挺好的，迈克尔。

I'm doing good, Michael.

很高兴能来。

Happy to be here.

还有朱莉·霍耶也一起来了。

And also joined by Julie Hoyer.

朱莉，欢迎你。

Julie, welcome.

你好。

Hello.

你最近怎么样？

How are you doing?

我很好。

I'm doing great.

我早就想聊这个话题了。

I cannot wait to talk about this topic.

我知道。

I know.

我也很期待。

I'm excited as well.

我是迈克尔·赫尔布林。

And I'm Michael Helbling.

所以这次节目，我们非得请一位嘉宾不可。

And so for this show, we absolutely needed a guest.

温斯顿·李是Arima公司的创始人，这家公司专注于合成数据和营销组合建模。

Winston Li is the founder of Arima, a startup specializing in synthetic data and marketing mix modelling.

在此之前，他曾在普华永道加拿大和奥美媒体集团领导数据科学团队。

Prior to that, he led data science teams at PwC Canada and Omnicom Media Group.

他还是东北大学的讲师，并担任该校数据分析硕士项目的顾问委员会成员。

He's also a lecturer at Northeastern University and sits on their program advisory committee for the masters in analytics.

今天，他就是我们的嘉宾。

And today, he is our guest.

欢迎来到节目，温斯顿。

Welcome to the show, Winston.

谢谢。

Thank you.

谢谢。

Thank you.

很高兴认识你。

Great to meet you.

很高兴认识大家。

Great to meet everyone.

谢谢您，迈克尔。

And thanks, Michael.

太好了。

Awesome.

关于合成数据这个话题，我觉得最好的切入点就是先说清楚它到底是什么。

Well, I think a great place to start on this topic of synthetic data is really just to talk about what it is.

您怎么定义合成数据呢？

So how would you define synthetic data?

它和匿名化数据或者抽样数据，从根本上有什么不同呢？

And what makes it fundamentally different from anonymized or sampled data, let's say?

是的。

Yeah.

非常好的问题。

Very good question.

简单来说，合成数据就是由算法生成的数据集，而不是从真实事件中收集来的。

So synthetic data, to put it simple, it's datasets that are generated by an algorithm as opposed to being collected from some sort of real event.

就我们做的人群数据领域而言，我们不会去搞问卷调查，不会追踪任何人，也不会要求任何人注册或提供信息，完全不是那回事。

So specific to consumer data, which is the space that we work in, we're not going out to conduct surveys, we're not tracking people, we're not asking people to provide anything to sign up, none of that.

合成数据的意思是，我们开发一个计算机算法，能生成看起来像真实数据的结果。

Synthetic data simply means we develop a computer algorithm where we could generate data in a way that mimics real data, so to say.

我们不是随便乱扔数字，而是根据特定的模式来生成数据。

So we're not just randomly spreading out numbers, we're generating data based on certain patterns.

这里有两个关键点需要注意。

Obviously, are two things that we should note.

第一，我们并不是在编造数据。

One, we're not making up data.

很多人一听到合成数据，就以为是假数据，觉得是捏造的。

A lot of people when they say synthetic data, they think of the word fake, they think it's fake data.

不是这样的，我们用来生成合成数据的算法，其实是用真实数据训练出来的。

It's not, And the algorithms that we use to generate synthetic data is indeed trained on real data.

所以，合成数据是基于对真实数据中模式的学习来生成的。

So it is based on learnings of patterns from real data in which we generate synthetic data.

首先，它不是假数据，而是非常接近真实数据，而且在统计分析中很有用。

So first point is that it's not fake, it's just like real data and it is very useful for statistical analysis.

关于隐私的问题，我们稍后会详细讨论，这也是合成数据最主要的应用动机。

We'll get to the whole discussion of privacy a little later on, which is the main motivation of synthetic data.

但从实用性来看，理论上它应该和真实数据一样有效。

But from a utility standpoint, in theory, it should be as useful as real data.

我还想指出一点：其实我们日常生活中早就接触过很多合成数据，只是我们没意识到，它们有别的名字，但本质上和我们说的合成数据是一回事。

The other thing I also want to point out is there's actually a lot of synthetic data in our day to day lives that we simply don't realize that being synthetic data, we call them something different, but they are in fact in along the same lines of same motivation, so to say.

比如你想想MidJourney这样的工具，它生成的就是合成图像。

If you imagine something like a mid journey, we're generating synthetic images.

如果把图像看作数据，那么像合成的人脸、不同地点、动物、风景这些图像，都是合成数据。

If you consider image to be data, then synthetically generated images like synthetic faces or synthetic pictures of different places, animals, sceneries, whatever.

这也是一种合成数据，本质上是根据学习到的模式生成虚拟像素，最终看起来就像一张真正的图片。

That's a form of synthetic data too, where it's simply generating synthetic pixels, so to say, but based on the patterns so that they look like a picture in the end.

ChatGPT 或 Gemini 也是同样的道理。

Same thing with even chat GPT or Gemini.

如果你把文字看作数据，那它们也是一种合成数据。

If you consider, again, if you consider words to be data, that too is a form of synthetic data.

用计算机基于真实信息的学习来生成某种信息，这种做法其实比我们意识到的要普遍得多，只是我们往往没意识到——从广义上讲，这就是合成数据。

So the fact that we use computers to generate some form of information, let's say, based on learnings from real information is much more common than we see and is much more common than we recognize and broadly speaking that that is synthetic data.

我以前完全认为合成数据是凭空捏造出来的，但说实话，几个月前我在纽约新泽西的 Measure Camp 上听了你的演讲，受益匪浅，对这个领域有了全新的认识。

I was 100%, I'll admit in the camp of thinking synthetic data just like was something you materialized out of thin air, but I have to say that I did have the benefit of getting to see you present at Measure Camp New YorkNew Jersey a couple of months ago, which was a fantastic presentation and I learned a ton about it.

我们刚聊起来，我想问问你，在我们这个领域，大家最常拿合成数据来解决哪些实际问题？

I think one of the other questions I love to ask you as we're kicking this off, it's like, what are some of the really common use cases that people within our field are using synthetic data?

它到底能解决哪些具体问题？

What problems is it really solving?

某种程度上，合成数据并没有创造出全新的应用场景。

Yeah, in some ways, synthetic data does not attain to new use cases, so to say.

并不是说合成数据能做些什么，而真实数据却做不到。

It's not like there is something synthetic data can do that real data cannot do.

人们更倾向于把合成数据看作一种手段，用来突破隐私法规的限制，做那些本来不能做的事。

People consider synthetic data more as a way to, let's say, be able to do things that privacy laws don't otherwise allow them to do.

某种程度上，人们用合成数据，是因为用真实数据来做同样的事，虽然技术上可行，但在采购或法律层面实在太难了。

So in some sense, people are doing certain things with synthetic data because to try to do the same thing with real data, while technically possible, is from a procurement or from a legal standpoint, very, very difficult.

其实合成数据本身没什么特别的，它就是这么回事。

So there's actually nothing special about synthetic data other than the fact that it is.

那我们是不是就到这里结束？

So should we just wrap it up here?

我不希望人们觉得合成数据和真实数据有本质区别，非得二选一。

I want people to think synthetic data is fundamentally different than real data in any way that I have to pick one or the other.

我想到最贴切的比喻，就是复印件。

The best analogy I can think of is, it think like a photocopy document if you will.

你有一份原件，但因为怕弄丢或者有其他顾虑，就不敢用它，于是复印了一份。

You have an original document, you don't want to use it, you're scared to lose it for various reasons, you make a photocopy of it.

这份复印件足以发挥和原件一样的作用，带来同样的实用价值，但用复印件更安全，你可以随便写写画画，不用担心弄丢它——在数据场景里也是这个道理。

That photocopied version will serve, will bring enough use, will bring enough utility just like the original document, however, it's much safer to work with the photocopy version because you can write on it, you're not afraid that you're going to lose it in a data scenario.

当然，你也不用担心会被人告上法庭。

Obviously, you don't have to worry about people suing you for doing various things.

不用害怕泄露别人的个人信息。

You don't have to worry about you know, revealing identity of somebody.

也不用担心有人选择退出。

You don't have to worry about people opting out.

所以，它让你在获得和真实数据同样效果的同时，还享有了所有的安全优势。

So it brings you all of the safety benefits while achieving essentially the same thing that real data will help you do.

大家好。

Hey, folks.

咱们来聊聊数据管道吧。

Let's talk about data pipelines.

你说有什么比可扩展的架构和安全的ETL更让人兴奋的呢？

You know, nothing says fun like scalable architecture and secure ETL.

对吧？

Right?

好吧。

Okay.

也许吧。

Maybe not.

但你知道什么才有趣吗？

But you know what is fun?

真正让人安心的是，不用整天担心敏感数据在云里乱飘。

Not actually sweating over whether your sensitive data is floating around the cloud uninvited.

这就是Fivetran混合部署的用武之地。

That's where Fivetran's hybrid deployment comes in.

它完美平衡了‘我们得完全掌控’和‘拜托，能不能有人帮我们管一下’这两个需求。

It's the magical middle ground between we need full control and please, somebody else manage this thing for me.

你可以在自己的环境中自主部署数据管道，同时享受Fivetran提供的便捷监控和自动更新。

You get self hosted pipelines in your own environment, but with all the ease of monitoring and updates handled by Fivetran.

这就像是你拥有一辆豪华座驾，而Fivetran的维修团队全天候为你维护保养。

It's like owning a fancy car and having Fivetran's pit crew keeping it running twenty four seven.

云部署在本地，还藏着一层秘密技术。

It's cloud based on premises, secret volcano layer.

这其实不重要。

It doesn't matter.

Fivetran提供了一个统一平台，全面管理这一切，内置的治理与安全功能让合规团队安心，工程师也能保持精力充沛。

Fivetran's got a unified platform that will manage it all with governance and security features that keep compliance teams calm and engineers caffeinated.

简单说，你的数据安全了，生活更轻松了，数据管道始终在线。

Bottom line, your data is secure, your life is simpler, your pipelines always on.

前往 fivetran.com/aph，立即开始为期14天的免费试用。

Go to five tran dot com slash a p h and start a fourteen day trial today.

就是 fivetran.com/aph。

That's fivetran.com/aph.

你尽管去吧。

Go ahead.

给你的数据应有的保障和支持。

Give your data the treatment it deserves.

不过我一直很好奇，要怎么保证这套东西能足够贴合真实的业务场景呢？

So my head always goes to how do you make sure it reflects the real world enough, though?

毕竟我记得你提过，这不是人们凭空编造出来的东西。

Because I know you said it's not something that people just make up.

它本来就是要基于实际情况来搭建的。

It's supposed to be based on the real world.

但我实在没法理解你说的这个应用场景，我之前和同事讨论过类似的事，当时那个提法就让我特别不舒服——我们那时候没把它叫合成数据。

But I can't wrap my head around the use case you're talking about because I've talked to colleagues about this in the past and I was really uncomfortable with this idea of we didn't call it synthetic data.

我们当时管这个叫对缺失数据建模。

We called it modeling missing data.

本质上都是处理那些没同意授权使用数据的用户的情况。

And it was all around people not opting in.

懂了。

Okay.

那如果他们选择了参与，这些数据会是什么样子？

Well, what would that traffic be doing if they had opted in?

我们能填补数据里的这些空白吗？

Can we fill those blanks in our datasets?

我脑子里一直想，不管对错，我真的搞不懂这怎么行得通。

And my head, right or wrong, I was like, I can't compute how this will work.

这怎么避免得了偏见呢？

How can it not be biased?

我们怎么知道那些根本不想参与的人，是怎么一步步选择退出的？

And how do we know what people who inherently don't wanna opt in are going through the steps to opt out?

也许他们的行为和我们能抽样建模的那些人完全不同。

Maybe they act completely different than the people we can sample to do the modeling on.

所以也许我们真该深入探究一下这个问题。

So maybe we wanna go down that rabbit hole.

你怎么才能让它不带偏见，或者至少尽量减少偏见？

Like, how how do you make it not bias or try not to?

对。

Yeah.

而且数据源本身也在变化，还会定期更新。

Well, data sources are also changing, are also being updated regularly.

所以构建合成数据集这件事，其实和开发软件很像。

So building a synthetic dataset is very much like writing software.

你不能写完代码就扔在那儿不管，代码放久了就会过时失效。

You don't just write code and kind of leave it there and code will go stale.

总得有人来维护这套东西。

Somebody has to maintain it.

数据集也是这个道理。

Same thing with the datasets.

时间一长，问题就会出现——首先就连样本组这类东西都有问题。

Over time, you're going to have So first of all, even things like panels.

这些样本组每年或者每半年就会更新一次，至少我们公司是这样的，我们会跟进最新的样本数据，重新运行、重新训练模型。

Every year they do it or every six months they have an update and you're going to pick, at least in our case, we pick that up and we rerun, retrain the models.

在某些情况下，数据源会随着时间推移变得过时，比如数据来源的声誉下降，或者其调查面板质量变差。

In some cases, a data source becomes stale over time, like maybe a data set source loses its reputation or their panel becomes bad.

我们还有一些其他的数据来源。

We also have data sources.

这在加拿大是个典型例子。

This is a case in Canada.

有个数据源为了省钱，转用了一个不太可靠的调查面板。

We have a source where they try to save money by going into a different panel that's less trustworthy.

他们的数据集因此逐渐恶化。

And their data sets deteriorate over time.

所以我们的工作不只是生成合成数据，还得确保数据原料本身靠谱。

So part of the work is not only to generate synthetic data, but also to source the ingredients right.

这有点像下厨。

It's kind of like cooking.

作为厨师，不仅要会做菜，还得知道去哪儿买、买什么食材。

Job as the chef is not only to cook the food, but also where to buy and what food to buy.

这同样是整个问题的一部分。

That's also part of the equation as well.

这完全说得通。

That makes total sense.

但这样一来，我反而觉得，我之前提到的关于缺失数据建模的特定场景，可能并不适合用合成数据，除非你相信你所追踪的群体足够代表那些你没追踪到的人。

But then it kind of makes me feel like the specific use case of missing data, modeling missing data that I was referring to still may not be then a good candidate for synthetic data unless you believe your assumption that, like, those that you're tracking are representative enough of those that you're not tracking.

我觉得你说得对，你可能需要其他类型的研究来验证这个假设是否成立。

And I think to your point, it sounds like you'd want some other type of research or something to hopefully validate whether that assumption would hold or not.

否则，你其实只是在增加对已知人群的数据量。

But otherwise, you're kinda going under the use case of, like, I'm gaining more data on the population I know.

你并没有真正填补那些完全没数据的未知人群。

You're not necessarily replacing the unknown population that you have really no data on.

总归有个问题，就是你能把结论外推到什么程度，以及外推到什么地步才是合理的。

There there is always how there's always sort of the the the consideration of how much can you extrapolate and how much is extrapolation is appropriate.

这不仅适用于我们开发的合成数据生成模型，也是我还在大学二三年级学统计时最早学到的一件事。

And this is not only the case with building synthetic data or the synthetic generation models we develop, but also as a this is one of the first things I learned as a statistician when I was in third year university or second year university.

教我们建回归模型的时候，可能会说，比如你用房屋面积来预测房价，

When they teach us to build regressions, they might say if you build a regression model, let's say, where you forecast the housing prices based on, let's say, the square footage of the house.

如果你的数据范围是1000到5000平方英尺，那你就不能让模型去预测一栋500万平方英尺的房子，

And if the range of your data was from, let's say, 1,000 square foot 5,000 square foot, if that was the range of your input data, then you cannot ask the model to forecast a house that's 5,000,000 square feet.

这超出了你的数据范围，属于过度外推了。

That's outside of your scope, and that's too much extrapolation.

所以用合成数据的时候，同样也要考虑：到底多少算是合理的外推，多少又过头了。

So the same kind of consideration here with synthetic data too is how much extrapolation is extrapolation.

当然了，我们建模型本来就是为了能外推，要是不能外推，那还要模型干嘛呢？

And obviously, the whole point of building models is we want extrapolate in some way, we didn't want to extrapolate, then there would be no need for a model.

比较合适的例子是，比如我们完全没关于亚利桑那州养宠物的数据，但其他州都有谁养狗、谁养猫的资料，

An appropriate scenario might be something like, let's say, if we didn't have any, let's say if we were to look at pet ownerships, we didn't have any data, let's say in the state of Arizona, but we had data from other states on what kind of people own dogs, what kind of people own cats.

而亚利桑那州的人口结构数据是有的，

And then for the state of Arizona, we have, let's say, the general population data.

这种外推就比较合理。

That's a pretty good extrapolation.

如果你大致了解普通人是什么样子，哪些人养狗，哪些人养猫，尤其是周边州的情况，就可以合理推测亚利桑那州的人也会有类似的行为。

If you generally know what people look like, what kind of people, dogs, what kind of people own cats, and especially for the surrounding states, you can make the assumption that people living in Arizona will likely behave the same way.

这种程度的推断是合理的。

And that sort of extrapolation is appropriate.

但如果你的数据是比如‘人在木星上能活多久’这种话题，而关于这个话题根本没有任何数据，那任何推断都是不靠谱的。

Whereas if your data was like, I don't know how long will somebody survive on Jupiter or something like that, whereas it's like absolutely zero data whatsoever on that particular topic, any extrapolation you're going to make is inappropriate.

所以，什么算合理、什么算不合理，得靠判断，绝对没有一个固定公式，能让你随便输几个数字，合成数据就自动出来了。

So there is a little bit of a judgment call as to what is appropriate, and there's definitely no sort of fixed formula to say, well, let's just plug these numbers in and here comes the synthetic data and then we're done.

根本不是这么回事。

It's definitely not that.

这中间确实需要持续维护，既要维护数据来源和方法论，也要不断更新你的判断和假设。

There is certainly a level of maintenance, both in maintaining the source of the data, in maintaining the methodology, but also in keeping your judgment up to date or keeping your assumptions up to date, so to say.

是的。

Yeah.

听起来确实如此。

It sounds like it.

没错。

Absolutely.

我有个问题，因为之前你稍微提过一点，温斯顿，说到合成数据在隐私保护方面很有用，特别是当涉及个人身份信息非常敏感的时候。

So one of the questions that I had because you you kinda teased this a little bit earlier, Winston, about that synthetic data is helpful in some of the context for privacy concerns or where personally identifiable information is really sensitive.

很多时候，人们的想法是，哦，我们把数据清洗一下，或者去标识化，或者做一下聚合就行了。

And a lot of times people, you know, the solution is like, oh, we'll scrub that data or we'll de identify it or we'll aggregate it.

但有时候，这样处理后的数据又达不到实际问题对数据的要求。

But sometimes then that isn't the type of data you need for the problem at hand either.

所以，合成数据是不是在某种程度上，正好平衡了这两者之间的需求？

So is synthetic data, like, walking this balance between those two in some way?

或者你能再展开说说，合成数据是怎么帮助我们应对越来越突出的个人身份信息问题的？

Or if you could just expound a little bit on that and talk about some of the ways that synthetic data helps overcome some of the challenges that we're seeing more and more with the PII.

当然。

For sure.

仅仅对数据做脱敏处理往往还不够。

Sanitizing the data often is not good enough.

而且这种情况越来越严重，因为立法总是跟不上技术的发展。

And it is increasingly so because lawmakers lag technology, so to say.

比如，某种新技术出现了，或者一种新的追踪方式被开发出来，数据就开始被收集，等人们发现的时候，才会惊呼：‘这太危险了。’

Like, a piece of technology becomes available or a new kind of way of tracking people becomes available, and then data sets get collected and then people look at it, people go like, Oh, this is dangerous.

等立法者意识到问题，往往已经过去好几个月了，这时候他们才去修订隐私法规，限制这类数据的使用。

The lawmaker realizes that many months later, and then they update what privacy laws are on using this kind of data and not being able to use this kind of data.

我举个例子吧。

So let me give you an example.

我们能接触到的一个数据集是移动轨迹数据。

One of the data sets we have access to is mobility data.

这种数据基本上就是通过手机的定位信息来追踪人的行踪。

Mobility data basically tracks people's location based on your lat alone.

你随身带着手机，各种应用都会在你授权的情况下收集你的位置信息。

And so you carry your phone, it tracks you through a number of apps.

它们有一整套应用生态，都会在你同意的前提下获取你的位置数据。

They have a whole portfolio of apps that allow you to share your location where you're asked to share your location.

一旦你同意了，它们就会每隔几秒、每分钟，或者每隔几分钟就向你发送一次定位请求，获取你精确的经纬度。

And once you agree to it, they will ping you every couple of seconds or every minute or every few minutes, and they will get your exact lat and long.

这种数据在市场上是有人卖的。

So this is a type of data that people sell on the market.

要拿到这个数据集其实并不难。

You can fairly easily acquire that dataset.

它被归类为个人身份信息，因为里面没有姓名。

It is considered PII in the sense that there's no names attached to it.

没有邮箱，也没有电话号码。

There's no email, there's no phone numbers.

你唯一能拿到的，是一个叫设备ID的东西，通常是IDFA，具体看你用的是iPhone还是安卓手机，但一般是个经过哈希处理的设备ID，单看这个ID你认不出是谁，但你可以把它导入广告投放平台，精准定位到这个人。

The only identification you get is what's called then sort of a device ID, which is usually IDFA or depending on whether you're an iOS phone or an Android phone, but usually a hashed device ID that you would not be able to identify the person, but you can stick it into a media buying platform and target that person.

所以，正因为它是经过哈希处理的，才被算作个人身份信息。

So that is considered PII because it's hashed.

但问题是，每隔几分钟，你就能收到一个人的定位信号，精确知道他当时在哪儿。

However, every couple of minutes, you're going to get a ping for somebody and you're going to find the exact lat and long of that person.

现在手机的GPS定位非常精准。

And the phone GPS is pretty accurate these days.

我想精度大概能到两三米左右。

I guess it goes down to maybe ten, twenty feet in accuracy.

你可以直接看某人的手机定位，把位置标在地图上，半夜三点钟也能知道他在哪儿，甚至能直接上门敲门，还能还原出他一整天去过哪些地方。

So you could totally look at somebody's phone, plot it on the map, figure out where they are, 3AM at night and then knock on their door and you know exactly where that person had been to in different parts of the day.

这说明光靠匿名化根本不够用。

So that is an example where just anonymization isn't really enough.

现在这事儿还处在灰色地带。

It's a little bit of a gray area now.

目前根本没有明确的政策规定，移动数据能怎么用，什么才算个人身份信息等等。

There's no policy on how mobility data can be used, what exactly is considered PII and so on.

如果你再把这套数据和人口普查数据结合起来，比如盯着一个特别小的区域——美国最小的人口普查区块里，人口还不到五十人。

And then if you imagine joining that dataset with, let's say something like the census, where you might be looking at a very small area, the smallest census block in The US has less than 50 people.

只要你掌握某人的设备ID，知道他住在那里，再结合不到五十人的普查数据，就能很容易推断出这个人很多个人信息。

So you could pretty easily identify somebody if you had someone's device knowing they lived there and then took the census with less than 50 people, you could very easily guess a lot of data about that person.

所以这只是其中一个例子，说明每个单独的数据源虽然都做了匿名处理，表面上看也符合隐私规范。

So that's just one example of how, where each of the individual data sources is anonymized and quote unquote privacy compliant.

但每一份数据其实都透露了关于这个人的某些信息。

And yet they each reveal something about that person.

当你把所有这些零散的信息拼在一起，就能对这个人有相当清晰的了解。

When you piece all these different information together, you get a pretty good idea of who that person is.

这就是为什么在大多数情况下，仅仅清洗数据是远远不够的。

This is why just sanitizing data isn't enough in most use cases.

正因如此，我们才尝试生成合成数据集，从根本上解决这个问题。

And this is the whole point why we sort of recreate synthetic copies of datasets to deal with that problem altogether.

如果非要说从我刚才讲的内容里能记住一点，那就是：如果可能的话，把手机上所有应用都卸载掉。

So if there's one takeaway from what I just said, that would be to uninstall all of the apps on your phone, if possible.

不然的话，你的行踪迟早会被追踪，信息也会被买来卖去，最终可能进入别人的数据库。

Otherwise, you're going to get tracked and your information is going to get bought and sold, and it might end up in people's databases.

其实我上个星期就把所有应用都关了，因为看到一个视频吓到我了，说他们掌握了你所有的位置信息。

So I turned all mine off the other week, actually, because I got some, like, scary, you know, reel telling me, like, they have all your locations.

听你这么说，我对自己当初那样做感觉好多了。

So hearing you say that makes me feel good that I did that one.

所以你的意思是，合成数据的一个应用场景，是通过加入合成数据来增加数据量，让它更具代表性？

So do you do you then are you saying a use case for Synthetic data could also be that by adding in Synthetic data, you're almost adding, like, representative, maybe, like, volume.

所以我不太想叫它噪音，但更像是精心设计过的干扰。

Therefore, I don't wanna call it noise, but it's almost, like, well designed noise.

这样能帮助提升多个数据集组合时的匿名性吗？

So would it help with the anonymity of, like, combinations of data sets?

我理解得对吗？

Am I understanding that?

其实不是噪音的问题。

It's not so much noise.

那是另一个完全不同的领域，完全不是一回事。

That's a different sort of area in No, data not round term at all.

那些做差分隐私的人，才是走的这条路。

Are people who do what's called differential privacy, which is actually along that line.

基本上，你拿一个数据集，加点噪声进去，别人反而可能反过来推断出这些数据的来源。

Basically, you take some dataset, you add some noise, you add some noise to the data, then people could go back and identify where that data comes from.

这是另一个方向，但我其实也在解决同一个问题。

So that is a different area, but I'm trying to address the same problem too.

对我们来说，合成数据更侧重于重新构建数据集，保留原有的统计特性，但具体的数值都变了。

For us, synthetic data is much more about recreating the data set in such a way that statistical properties are preserved, but the actual sort of cells are different, so to say.

举个例子，假如我只关心平均值，有三个数，比如一、三、五，它们的平均值是三。

So one example of that would be, let's say if the only thing I cared about was average, I have three numbers, let's say, one, three, five, as an example, and their average in this case is three.

那合成后的数据可能是二、三、四。

A synthetic recreation of that might be something like two, three, four.

数字不一样了，但平均值还是三。

Different numbers, but the average is also three.

所以，一、三、五和二、三、四，是两组不同的数字。

So in this case, one, three, five and two, three, four are two sets of different numbers.

但就算平均值而言，这两组数的作用完全一样，算出来的结果一模一样。

But for the purposes of calculating the average, these two numbers serve the exact same purpose because they give you the exact same average.

当然，实际情况更复杂，因为统计学家和数据科学家会关注更多指标。

Now, of course, it's more complicated than that because statisticians or data scientists look for many more metrics.

他们会看均值、相关性，还有各种各样的其他参数。

They'll look for mean, they'll look for average, they'll look for correlation.

要满足的条件一大堆，你的合成数据得把所有这些都通过才行。

There's a whole bunch of check boxes and your synthetic data has to tick all of the boxes.

但这个例子只是想说明，我根本没有添加任何噪声。

But that's just a simple example of saying, I'm not adding any noise.

我的新数据集可以是2、3、4，也可以是4、3、2，或者4、2、3，顺序无所谓，我只关心最终的统计结果。

My number could be, my new dataset could be two thirty four, it could be four thirty two, it could be four twenty three, doesn't matter what order it is, but all I care about is the statistical metric in the end.

这个结果不会变。

That doesn't change.

这才是合成数据真正要做的事。

That's that's really what synthetic data is.

你也可以把那个小镇的50这个数字调高一点，这样就更难锁定到具体某个人了。

But you could beef up the 50 in that little town in hopes that, you know, it'd be harder to pin it down to one person.

是的

Yeah.

太棒了

Awesome.

没错

Exactly.

有一件事，我特别高兴你问了这个问题，朱莉娅。

One of the things and I that was a really good I'm so glad you asked that question, Julia.

你解释得非常好。

It was a really good explanation.

温斯顿，我在你的一篇论文里看到过‘辅助低分辨率数据’这个说法，我觉得这个表述特别贴切，因为在我看来，这正好呼应了朱莉娅说的内容——通过稍微扩大样本量来帮助解决匿名性问题。

One of the things I saw in one of your papers, Winston, was the phrase assisting with low resolution data, which I just that framing really connected with me because it my interpretation of that is it helps with exactly what Julia was talking about, assisting by kind of, like, increasing that small base size a little bit and in dealing with some of the anonymity.

而且，当数据被聚合到完全失去意义的时候，它也能派上用场。

But also, it helps when, like, the aggregation it's aggregated it to the point where it's no longer helpful.

对吧？

Right?

所以它需要提供更细致的粒度，但不能让人识别出具体个人。

So it needs to provide more granularity, but not in a way that's identifiable.

这是否就是这句话的本意？或者在其他情况下，人们是怎么解决这种低分辨率问题的呢？

Is that, like, what is intended by that, or what are some some ways that the low resolution kind of problem, if you will, is kind of solved in some other cases?

是的。

Yeah.

要回答这个问题，我先简单说说我们在谈数据集时所说的‘分辨率’是什么意思。

To answer that question, let me maybe just discuss a little bit about what we mean by resolution when we talk about data sets.

在消费者数据中，人们经常把数据做聚合处理。

In consumer data, it's very common for people to aggregate data sets.

所谓聚合，就是你手上有大量个体的数据，但出于隐私考虑，不能泄露任何一个人的具体信息。

What we mean by that is you have lots of individuals, but because of, again, privacy, you can't give away information about each of the individuals.

不过我们可以把他们分组，归入不同的群体，然后给出平均值。

But we can group them, we can put them into cohorts and give you averages.

而最常见的做法就是按地理位置来划分。

And a common way of doing that is by using geography.

比如在美国，这可能是人口普查区块、邮政编码、县，或者设计市场区域，随便哪种都行。

So in The US for example, that could be census blocks, that could be a zip code, that could be a county, that could be a DMA, whatever.

对吧？

Right?

我们不会说鲍勃年薪五万，莎拉年薪三万，而是说这个邮政编码区域的平均收入是十万，住着一万人，这就是平均值。

So rather than saying, Bob makes $50,000 a year, Sarah makes $30,000 a year, rather than saying that, we say, well, this zip code, the average income is $100,000 There are 10,000 people living in that zip, that's the average income.

另一个邮政编码区域，平均收入可能是三万。

There's other zip, average income is $30,000 for example.

用地理区域来聚合数据有很多好处，部分是因为地理边界清晰明确，容易界定。

And there are many advantages of using geography as a way of aggregating data, partly because geography is quite well defined and straightforward.

人人都知道自己住哪个邮编，属于哪个区域。

Everybody knows what zip they live in and which geography they belong to.

但最重要的是，地理区域不会抱怨。

But most importantly, geographies don't complain.

邮政编码不会说：‘你侵犯了我的隐私。’

Zip code will not complain and say, oh, infringed my privacy.

你得把我从这里面删掉。

You need to take me off of that.

把我从你们的数据集里移除吧，随便你怎么叫。

Take me off of your dataset or whatever.

所以用地理区域来汇总数据，这是很常见的方式。

So it's a very common way for people to aggregate data.

不过，用地理汇总数据的问题在于，该选哪个层级的区域。

Now, the issue with aggregating geography is what geography to use.

因为如果区域太小，比如假设每两户人家算作一个地理单元，那意义就不大了——我和隔壁邻居被分到同一组。

Because if the geography is too small, like if you're looking at, hypothetically, let's say if every other house was grouped into some kind of a geography, that's probably not very helpful because I and my neighbor are in the same group.

要是谈平均收入，我知道自己的收入，马上就能推算出他或她家的收入情况。

If we're talking about average income, I know mine, I can immediately calculate what his or what her average what her family income is.

这就是地理单元太小的后果，个人还是能推断出别人的具体信息。

So that's an example of where your geography is too small and individuals could infer what people still could infer personal information or personal data.

反过来，如果地理单元太大，比如整个国家算一个区域，那平均的人太多，也没多大用处。

And then on the other hand, you have a problem where if your geography is too big, let's say if your geography was the whole country, you're averaging too many people and that's not very helpful.

所以如果有人说，美国人的平均年龄，比如说住在美國的人平均是48岁。

So if people say, well, the average age of The United States is, let's say people living in The United States is, let's say 48 years old.

那又怎样？

Well, what?

对其他国家来说，大概也是这样，因为你把太多人混在一起平均了。

That's probably true for every other country out there, because you're simply averaging too many people.

人一多，差异就被稀释了，平均值也就变得没那么有意义了。

When you average too many people, the variation sort of washes out and you just well, up the average.

所以找到合适的平衡点很重要。

So finding the right balance is important.

我们说的高分辨率，指的是数据更贴近个人层面，更细致；低分辨率呢，就是很多人被合并、平均了，信息在过程中就丢失了。

And that's what we mean by, when we say high resolution, we mean the data is more granular at the closer to the individual level, low resolution, meaning a lot of them have been grouped up, averaged, and you lost information as a process.

在美国，人们通常会选用普查区块作为基本单位。

In The US people typically, census block is a pretty common choice.

邮政编码也是一个不错的选择。

Zip is a pretty good choice as well.

所以again，具体用哪个地理单元，得看组织本身，还有他们能接受的风险程度。

So again, depending on the organization, depending on what their risk tolerance is, those are usually the geographic units that people work with.

那这么说，回到瓦尔的问题，低分辨率其实并不符合我们之前说的那种情况——一个小普查区块里有50个人，再和其他数据集合并，反而更容易识别出个人。

So then I guess to loop back to Val's question, low resolution would not be describing the scenario we were talking about where there's 50 people in a small census block and you're combining it with another dataset where it becomes somewhat more identifiable.

那我们要不要回头再问问瓦尔？

So I guess do we wanna circle back to Val?

你之前提到，低分辨率的问题上，合成数据能帮上大忙，对吧？

Because you were kinda asking then for the low resolution problem, you said synthetic data can help a lot there.

没错。

Right?

那举个例子，用低分辨率、高聚合的数据时，合成数据具体是怎么起作用的？

So what would an example, I guess, be where synthetic data helps when you're using more aggregated low resolution datasets?

当你处理低分辨率数据集时，通常会面临一个权衡：高分辨率数据信息更丰富、更细致，但隐私风险也更高。

Well, when when you're dealing with low resolution data sets, what typically happens is So there's a little bit of a trade off, high resolution data, more information, more granular, but less safe from a privacy standpoint.

反过来，低分辨率数据经过更多聚合，细节少了，信息也丢了，但正因如此，个人身份反而更难被还原出来。

And then on the other side where you have low resolution data, that's more aggregated, less granular information being lost in the process, but information being lost also means identities are lost and it's harder to connect back to the individual.

所以尤其是在处理高分辨率或低分辨率数据集时，合成数据能有效补充信息量。

So especially working with high resolution, low resolution data sets, synthetic data could be a very helpful augmentation to getting more information.

如果你手头已经有低分辨率数据，或者个体级别的数据，那就根本不需要合成数据。

If you have low resolution data, if you have individual level data, then you don't need synthetic data.

你已经有需要的一切了，直接用就行。

You have what you need, so you can just go do whatever you need to do.

用合成数据的核心意义在于，当你没有高分辨率的精细数据时，它能帮你以统计上可靠的方式达成目标。

The whole point of leveraging synthetic data is when that high resolution granular data is not available, and then synthetic data becomes your alternative to do, to get to what you need to do, at least in a statistically meaningful way.

当然，合成数据集未必和真实高分辨率数据完全一致，但如果你用它来建模或分析，效果至少能和用真实数据差不多。

Not to say your dataset necessarily matches exactly the same as the low resolution data if you were to have it, But at least if you were to build models or do analysis on this synthetic data set, you can expect it to work just as well as if you were to use the real high resolution

数据，哦，明白了。

data Oh, okay.

不好意思，我刚才打断了。

This may sorry.

我感觉自己问的问题可能挺基础的，但对我个人来说，这些信息真的特别有帮助。

I feel like I'm asking really naive questions, but this is so helpful selfishly for me.

所以当你用合成数据处理低分辨率数据时，是为了增加低分辨率数据的样本数量，也就是更多的数据行吗？

So when you're doing synthetic data for low resolution data, are you doing it to get more of the low resolution, like, data points or rows of data, whatever it is?

还是说你用合成数据是为了生成一种不涉及真实个人的、更高分辨率的虚拟数据源，来补充你的低分辨率数据？

Or are you saying you're doing synthetic data to create a synthetic safe, it's not real people, higher resolution data source to augment your low resolution?

或者两种方式其实都可以结合使用？

Or maybe you could do both?

是第二种。

The latter.

更准确地说，是第二种。

More more the the latter.

明白了。

So Okay.

我通常这么理解：这有点像数据压缩，你原本有高分辨率的数据。

The way I describe it is it's like a little bit of it's it's a it's like data compression where you have high resolution data.

比如说，你原本有个人层面的数据，然后把这些数据聚合到邮编或人口普查区块这样的层级。

Let's say you have individual level data and people aggregated into, let's say, zip codes or census blocks.

把人数据聚合起来，是为了保护隐私。

So people aggregated to preserve privacy.

那我们现在能不能用这些聚合后的数据，通过建模或算法，重新还原出被聚合的个体信息？

And then now can we take the aggregated data set, apply some modeling, apply some algorithm to reconstruct the individuals that has been aggregated?

这就像是数据的压缩和解压缩，但目的不是为了节省硬盘空间，不像把一堆文件打包成ZIP文件那样。

So it's like data compression and decompression, except you're not compressing data to save disk space, like how you put a bunch of files into a zip folder.

我们不是为了省空间，而是通过这个压缩和解压的过程，把身份信息去掉了。

You're not doing it to save disk space, but from that compression and decompression process, you lost, you took away the identity.

天哪，我明白了。

Oh, oh my gosh.

现在我总算搞清楚了。

Okay, it's clicking so much more.

谢谢。

Thank you.

对我而言，这个解释真的很有帮助。

For me too, this is really helpful.

我觉得另一点是，从数字营销的角度来看，当你想定位和筛选目标受众时。

And I think one of the other things, and I think about this from like, a digital marketing perspective, like, when you're trying to identify and pick audiences.

比如你知道有多少人养宠物，就像你之前提到的那样。

So you know how many people have, like, the pets, like, you were talking about before.

现在你还想弄清楚，其中有多少人对高等教育项目感兴趣。

And now you wanna know also how many of them are interested in, you know, some higher education program.

所以你得搞清楚，如果你要同时满足多个条件——比如养宠物、又对教育感兴趣——你的目标人群到底有多大。

So you have to understand or try to figure out what is the size of my audience if I'm picking, like, has this, this is true, also this, this.

但这些交叉组合很难估算，因为你手头只有类似‘这个邮编区有30%的人养狗’这样的数据。

And so those combinations aren't easy to come by to try to, like, estimate how many people fall into that bucket because all you have is, you know, thirty percent of people have dogs in this, you know, zip code.

对吧？

Right?

所以这就像对的那样。

So is that kinda like Yeah.

你通过生成虚拟人群来提升估算的准确性，让统计数据更接近真实情况——你之前是怎么说的来着？

Where the where you get better with your estimations or audience sizes because you're making fake people that where, like, the statistics are still what was what how did you phrase it earlier?

就是说

Where, like

这个

The

我忘了你之前怎么称呼那种方法了，但你得确保它仍然有效。

I forget what you called it earlier that you would make sure it's still.

对，从统计上来说是等价的。

Yeah, yeah, statistically equivalent.

对，没错，没错。

Yes, yes, yes.

好的，明白了。

Okay, yes.

对，就是这样。

Yeah, exactly, exactly it.

我给你举个简单的例子。

So let me give you a simple scenario.

假设一个社区里有十个人。

Let's say in a neighborhood we have 10 people.

根据人口普查数据，假设其中一半是男性，一半是女性。

According to the census, let's say half of them are male, half of them are female.

同样根据人口普查，一半的人养狗，另一半养猫。

And according to the census again, half of them own dogs, half of them own cats.

那现在你问，这个社区里有多少男性养狗？

And then now you ask, how many males own dogs in that neighborhood?

光看这些汇总数据，你是没法知道的，根本无从判断。

Well, just by looking at the aggregated information, you don't know, you won't know.

可能一种极端情况是，五个男性全养狗，五个女性全养猫。

You could have one extreme where all of the five males own dogs and all of the females own cats.

或者反过来，五个男性全养猫，五个女性全养狗，也可能两种情况混在一起，分布很随机。

Or you could have it flipped, all of the five males own dogs, all of the five females own cats, or you could have a pretty random mix of the two.

不管哪种情况，总的数字看起来都一样，但实际含义却天差地别。

And in all cases, your aggregated counts are the same, but they would mean very, very different things.