In Hollywood, it is known that the sequels are rarely better than the original movie/parts. Batman: The Dark Knight Trilogy is a notorious exception. I believe that this post is yet another.

在好莱坞，众所周知，续集很少比原始电影/部分更好。 蝙蝠侠：黑暗骑士三部曲是一个臭名昭著的例外。 我相信这是另一篇文章。

After explaining what many claim to be a “Full-Stack Data Scientist” in Part I of this post, I am going to discuss the main claims that sustain “the generalist” as a way to go for a Data Scientists (DS) in industry. In this post, I present, discuss and deconstruct the four key arguments that “the generalist” fans use to support their point of view.

在解释了这篇文章的第一部分中许多人声称是“全栈数据科学家”之后，我将讨论维持“通才”的主要主张，以此作为行业中数据科学家(DS)的一种方法。 在这篇文章中，我介绍，讨论和解构“通才”支持者用来支持他们观点的四个关键论点。

“全栈数据科学家”的四大错误Struts (The four false pillars of the “full-stack data scientist”)

(I)根本原因分析是端到端的-因此，DS也应该如此！((I) Root Cause Analysis is end-to-end — so, a DS also should be!)

Root Cause Analysis is something strategy consultants have been doing for decades. One of the main pro-generalist argument is that the causes for the low performance of Machine Learning models lies outside the modelling stage. And, by experience, I can classify that statement as truthful. Now, two questions arise: (1) should RCA be done by data scientists? And, if yes, (2) would it help them to be more effective on finding the root causes if they own the entire process? And, so far, I still follow with them: yes and yes for the two questions. But quickly, another question arises: how would it be if data scientists owned this project end-to-end? Well, the most likely answer would involve delays on project completion, lower user acceptance, troubles on scalability…i.e., more root causes to find and a bigger problem. Why? Because they are not experts (e.g.: in UX design, devops or data engineering)! When you try to do something that is difficult and you are not an expert on it, you will take more time to do it and you will be more likely to do a mistake…or several. Ab uno disce omnes.

根本原因分析是战略顾问数十年来一直在做的事情。 普遍主义者的主要论据之一是，机器学习模型性能低下的原因不在建模阶段。 而且，根据经验，我可以将该陈述归类为“真实” 。 现在，出现两个问题：(1)RCA是否应由数据科学家完成？ 并且，如果是，(2)如果他们拥有整个流程，是否可以帮助他们更有效地找到根本原因？ 而且，到目前为止，我仍然跟随他们：对两个问题，是和是。 但是很快出现了另一个问题：如果数据科学家端到端拥有该项目，情况会如何？ 好吧，最可能的答案将涉及项目完成的延迟，用户接受度的降低，可伸缩性的问题……即，更多的根本原因被发现和更大的问题。 为什么？ 因为他们不是专家(例如：UX设计，devop或数据工程专家)！ 当您尝试做一些困难的事情而又不是专家时，您将花费更多的时间来做，而且您更有可能犯错误，甚至是犯几个错误。 不可否认的全能。

What is the problem that usually happens when specialized DS try to run RCA? They stumble onto other team’s walls. Why? Three classics are: lack of ownership ,aka Social Loafing, lack of communication and lack of data driven awareness. If other teams are not available to step up or their technical/communication mistakes and/or are not aware of what the data science modules do (input/output) and of the possible consequences that their (bad) work may have there…your organization has a problem of Culture, Values and, ultimately, Leadership. And, if that happens, you can have whatever type of Data Scientists you want…as you will always be closer to failure than to… well, pretty much anything else.

专用DS尝试运行RCA时通常会发生什么问题？ 他们跌倒在另一支球队的墙上。 为什么？ 三个经典是：缺乏所有权，又名社交闲逛，缺乏沟通和缺乏数据驱动的意识。 如果其他团队无法加强工作或他们的技术/通信错误，并且/或者不知道数据科学模块的工作(输入/输出)以及他们的(不良)工作可能带来的后果……您的组织有文化，价值观和最终领导力的问题。 而且，如果发生这种情况，您可以拥有所需的任何类型的数据科学家……因为您总是比失败更接近失败……嗯，几乎还有其他任何事情。

(II)多个角色带来沟通负担 ((II) Multiple Roles bring Communication Overhead)

I have to admit this is the one that makes less sense for me. If you work to develop a project in a multidisciplinary team, you need to communicate. If you work alone, you do not need to communicate….but, to stamp communication as overhead is something that goes beyond my comprehension. Some — well-intentioned individuals, I’m sure — mention that Data Science is an immature field and, consequently, it is difficult to create the blueprint of the final product beforehand — as we typically have in Software Engineering. That is something that I can even agree upon. Consequently, rapid (re)build-try-fail iteration is key for success. Naturally, that speed boost would come by full project ownership — aka taking the generalist way. And, oops…I found a fallacy here.

我必须承认，这对我来说意义不大。 如果您要在多学科团队中开发项目，则需要进行沟通。 如果您一个人工作，则无需沟通……但是，要加紧沟通，因为开销超出了我的理解范围。 某些人-我很有把握的个人-提到数据科学是一个不成熟的领域，因此，像我们通常在软件工程中通常那样，很难事先创建最终产品的蓝图。 我什至可以同意这一点。 因此，快速(重新)构建-尝试-失败迭代是成功的关键。 当然，这种提速将由项目的完全所有权(也就是采取通才的方式)来实现。 而且，糟糕…我在这里发现了一个谬论。

In software engineering, you generally have already multiple teams in place organized either in tribes, technologies or functions. They work together to accomplish a singular goal…but, on that process, they have competing agendas themselves…consequently, alignment and communication is key for their success. Naturally, impose such amount of meetings, stand-ups and other rituals to all staff level technical team members may be tough…and thus, also naturally, the role of the Agile Product Owner got born.

在软件工程中，你一般都已经到位多个团队在任何组织的部落，技术或功能。 他们共同努力以实现一个单一目标……但是，在这个过程中，他们自己拥有相互竞争的议程……因此，协调一致和沟通对于他们的成功至关重要。 自然，向全体员工级别的技术团队成员施加如此大量的会议，站立式会议和其他仪式可能很困难……因此，自然而然地，敏捷产品负责人的作用就此诞生。

Data Science is no different. And, although this is a recent trend, it is a popular one. To have a real impact in an organization, every Data Science team needs to have an (Agile) Product Owner in place (check its description below). And please…don’t tell me that your team does not have one and/or your organization is too small for that…because then, perhaps, that (the lack of a Product Owner for Data Science and/or a Team Leader/Head) is where the real problem started in the first place.

数据科学也是如此。 并且，尽管这是最近的趋势，但它是一种流行的趋势。 为了对组织产生真正的影响，每个数据科学团队都需要有一个(敏捷)产品负责人(请查看下面的说明)。 并且请……不要告诉我您的团队没有人，并且/或者您的组织太小了……因为那也许是因为(缺少数据科学的产品负责人和/或团队负责人/负责人)是真正的问题首先出现的地方。

(三)边做边学是关键！ ((III) Learning by doing is the key!)

Nowadays, there is a plethora of online courses for Machine Learning and Data Science following the early success of Coursera. Not rarely, these courses come with a promise like “Zero-to-hero: How to become a data scientist in 6 months” (I even found a blog post with that expression). In these courses, it is often said that Data Science is an empirical discipline where to learn by doing is key. Like in most of other jobs, I agree that experience on actually doing stuff (vs. only theorizing about it) helps. A lot. The problem is that expression implies that everybody can become a Data Scientist if they try enough. Without reducing the merit to people that studied other STEM disciplines (like Physics or Production Engineering) and then, with a lot of effort and persistence, made a difficult transition to Data Science careers…it is a fact that they will struggle to become Professional Data Scientists as they lack foundations about the topic. In other words, they know when and how to use the available off-the-shelf tools/libraries/methods till a certain but they don’t know exactly how it works. Please read more about my thoughts on Professional vs. Citizen Data Scientists on this keynote that I gave last year on IT Arena 2019 (Lviv, Ukraine), where I shared the stage with the likes of Microsoft, Spotify or Uber, among others.

如今，随着Coursera的早期成功，有很多关于机器学习和数据科学的在线课程。 这些课程很少能像“零到英雄：如何在6个月内成为一名数据科学家”这样的承诺(我什至找到了一个表达这种表情的博客文章)。 在这些课程中，通常会说数据科学是一门经验性学科，而边干边学是关键。 像在其他大多数工作中一样，我同意在实际做事方面的经验(相对于仅将其理论化)会有所帮助。 很多。 问题在于这种表达意味着每个人只要努力就可以成为数据科学家。 在不降低学习其他STEM学科(例如物理或生产工程)的人的才能的情况下，然后通过大量的努力和毅力，他们很难过渡到数据科学事业……这是他们将很难成为专业数据的事实科学家们缺乏有关该主题的基础。 换句话说，他们知道何时以及如何使用现有的现成工具/库/方法，直到一定程度，但他们并不确切知道它是如何工作的。 去年我在IT Arena 2019(乌克兰利沃夫)上发表的主题演讲中，请阅读有关我对专业数据与公民数据科学家的看法的更多信息，在这里我与Microsoft，Spotify或Uber等共享了舞台。

What are the consequences of such ignorance? Firstly, it prevents to really tailor their methodology to their business application (a classical step in CRISP-DM methodology). This entails that the result will not have so much impact in your business. This makes leaders wonder how they can extract more value from these Data Scientists who are delivering below expectations…and the answer invariably comes on about “getting them more stuff to do”. And that stuff would be devops, data engineering or analytics dashboards…does this ring a bell? Of course…it is a Full-Stack Data Scientist! :)

这种无知的后果是什么？ 首先，它阻止了根据他们的业务应用真正地调整其方法( CRISP-DM方法中的经典步骤)。 这意味着结果不会对您的业务产生太大影响。 这使领导者想知道如何从交付低于预期的数据科学家那里获取更多价值……而答案总是来自“让他们做更多的事情”。 而且这些东西将是devops ，数据工程或分析仪表板……这能敲响钟声吗？ 当然……它是全栈数据科学家！ :)

“Any fool can know. The point is to understand.” Albert Einstein

“任何傻瓜都知道。 重点是要了解。” 艾尔伯特爱因斯坦

I am doing hiring for positions in the all-data-things space (date engineers, data scientist — different flavors and machine learning researchers) for the last 4 years. For the latter two roles, my interviews always contains a technical Q&A on their past data science projects which covers, among other things, foundations of the used methods. Let me share with you some of the pearls that I’ve heard during these interviews in the last years:

在过去的4年中，我正在招聘所有数据事物领域的职位(日期工程师，数据科学家-不同口味和机器学习研究人员)。 对于后两个角色，我的访谈总是包含对他们过去的数据科学项目的技术问答，其中包括所用方法的基础。 让我与您分享我在最近几年的访谈中听到的一些珍珠：

Models trained with RandomForests get better as larger is the number of trees;

随着树木数量的增加，用RandomForests训练的模型会变得更好。

The difference between Logistic and Linear Regression is that one provides linear models…and the other doesn’t;

Logistic回归和线性回归之间的区别在于，一个提供线性模型……而另一个不提供线性模型。

(Answer to: why you are using Huber Loss when your evaluation metric is RMSE?) In Data Science, we need to try everything first to see what works the best. (He literally tried all the linear regression methods available on sci-kit learn).

(答案：为什么当您的评估指标为RMSE时为什么要使用Huber损失？)在数据科学中，我们需要首先尝试一切，以找出最有效的方法。 (他实际上尝试了sci-kit学习中可用的所有线性回归方法)。

AUROC stands for…Accuracy.

AUROC代表……精度。

(On comparing SVM’s against other learners…) The used model hyperparameters are the default ones. This is a fair comparison as the performance uplift after tuning those would be always small.

(在将SVM与其他学习者进行比较时…)所使用的模型超参数是默认参数。 这是一个比较合理的比较，因为调整后的性能提升始终很小。

(When asked what to do to create a model from a 500 examples training set) Use a CNN in Keras. Always. Preferentially, a pre-trained one.

(当被问到如何从500个示例训练集中创建模型时，使用Keras中的CNN。 总是。 最好是经过预训练的。

If you did not find something wrong with any of the above sentences, let me tell you right away: data science may not be your best career path.

如果您没有发现上述任何句子有问题，请立即告诉我：数据科学可能不是您最好的职业道路。

Work hard is mandatory…but to work hard & smart is even better!

努力工作是强制性的……但是努力工作和精干甚至更好！

Following this line of thought, another argument is that to own all the development process brings a higher sense of satisfaction. I am already imagining the joy on people’s face saying “I did it, I did it…it took me one month without weekends but I did it!”. Probably, then, the same full-stack data scientists will spend another month without weekends to fix the bugs that will be discovered after going live…but that is ok as he/she made impact :) On a more serious note…there is nothing wrong about working hard…that is key to any career, especially Data Science! But to work smart must be equally important. Investing time on doing things which you are strong at just makes you…stronger! If you doubt it, check out how seriously Cristiano Ronaldo trains scoring goals…but he certainly does not practice much goalkeeping :)

按照这种思路，另一个论点是拥有所有的开发过程会带来更高的满足感。 我已经在想象人们脸上的喜悦，说：“我做到了，我做到了……我花了一个月没有周末，但我做到了！” 那么，可能是相同的全栈数据科学家将花费另外一个月而不用周末来修复将在上线之后发现的错误……但这没关系，因为他/她产生了影响:)更严重的是……没有什么努力工作是错误的……这是任何职业的关键，尤其是数据科学！ 但是聪明地工作必须同样重要。 花时间去做自己擅长的事情只会使自己……更强！ 如果您对此表示怀疑，请查看克里斯蒂亚诺·罗纳尔多(Cristiano Ronaldo)训练得分目标的认真程度……但他当然不会练习太多的门将:)

(IV)准确性的-1％不会影响我的业务… ((IV) +-1% of accuracy does not impact my business…)

It is often said that +1% is not a performance uplift that is worthy to invest time into. It does not move the needle. Consequently, what you need actually is some guys (generalists) to use some data to build any model that works. That can actually be achieved on a quick-and-dirty way with some build-try-fail iterations and a lot of copy-paste from online helpboards such as StackOverflow. Make a MVP to run a POC. Have a spike…a proof point to show our investors that “yes, we can!”. All good fellas, aren’t they?

人们常说+ 1％的提升并不是值得花时间投入的。 它不会移动针。 因此，你实际上需要的是一些球员(通才)用一些数据来建立任何模型作品。 这实际上可以在一个快速和肮脏的方式取得了一定的积累尝试-失败的迭代，并从网上helpboards如大量复制粘贴的StackOverflow 。 制作MVP以运行POC。 突飞猛进……向我们的投资者证明“是的，我们可以！”的证据。 所有的好家伙，不是吗？

Do you have a data strategy & governance roadmap? No? Well, that’s bad…

您是否有数据战略和治理路线图？ 没有？ 好吧，那很糟糕……

Today, one of the problems of organizations is the lack of a data strategy/ roadmap and/or data governance policies. Often, senior leadership (both in corps and/or start-ups) is not aware of what creating and scaling a data-driven business can actually mean in practice. They have a business plan, sure, and naturally they like the scalability vs. reduced OPEX that automation brings into that…and of course, the competitive advantage that doing that fast can bring to their business. But they often forget to see the darker side of it. I like to call it to become data hostage.

如今，组织的问题之一是缺乏数据策略/路线图和/或数据治理策略。 通常，高级领导(无论是公司还是初创公司)都没有意识到在实践中创建和扩展数据驱动型业务实际上意味着什么。 当然，他们有一个业务计划，自然地，他们喜欢自动化带来的可扩展性与降低的OPEX……当然，快做可以给他们的业务带来竞争优势。 但是他们经常忘记看到它的阴暗面。 我喜欢称它为数据人质。

Copy-paste code: 1$. To know which code to copy-paste: 10000$.

复制粘贴代码：1 $。 要知道要复制粘贴的代码：10000 $。

There are a series of implications on automating your business in a data-driven way: you become dependent on the used data — which means that, if there is a bankruptcy of a data provider and/or a regulatory change that forbids of using that data in the future…well, you are screwed. You also become dependent on the pipeline dependencies — packages, development language, versions, etc. — even if they run in any type of virtual containers. Finally, you also become dependent on your model: not rarely, a model performance may change with the scale (accept 20% of customers instead of 5%, recommend 10 products instead of 3 and so on). Here, you expect to maintain your business performance (driven by your model’s performance) after scaling up…or even improve it. But, let me tell you the following secret: it is not any model that is capable of that. However, by then is too late: you already have a team of citizen data scientists, you already promoted the guys for their great work (after all, you did the POC, you got the new funding round, you convinced the investors) and you already have a series of processes (dependencies) of a group of people that is far from being specialized on the tasks they need to do….now, at scale.

以数据驱动的方式使您的业务自动化有一系列含义：您变得依赖于使用的数据-这意味着，如果数据提供者破产和/或法规变更禁止使用该数据将来……好吧，你被搞砸了。 即使它们在任何类型的虚拟容器中运行，您也将依赖于管道的依赖性(程序包，开发语言，版本等)。 最后，您还依赖于模型：很少会出现模型性能随规模变化的情况(接受20％的客户而不是5％的客户，推荐10种产品而不是3种，依此类推)。 在这里，您希望在扩大规模甚至改善它之后保持业务绩效(由模型的绩效驱动)。 但是，让我告诉您以下秘密：并非任何模型都能做到这一点。 但是，那时为时已晚：您已经拥有一个由公民数据科学家组成的团队，您已经为他们的出色工作提拔了这些人(毕竟，您完成了POC，获得了新一轮融资，使投资者信服了)，并且您已经有一群人的一系列流程(依赖项)，而不仅仅是专注于他们需要做的任务……现在已经大规模地进行了。

Do you want to scale your business one day? If yes, you need professional DS now.

您想一天扩展业务吗？ 如果是的话，你现在需要专业的DS。

Let me tell you another secret: if your business has (or want to have one day) scale and to have a predictive or prescriptive analytics engine in its core…you need professional DSs now. Probably, a series of other roles as well…but those, at least, for sure. And if not…well, perhaps your a company is not a place where a DS should aim to work at.

让我告诉您另一个秘密：如果您的企业具有(或希望有一天)规模，并且在其核心中具有预测性或规范性分析引擎……您现在需要专业的DS。 可能还有一系列其他角色……但至少可以肯定的是。 如果不是……那么，也许您的公司不是DS应该致力于的地方。

If you have such automated data-driven business in place, the impact of +1% in your model’s performance is tremendous…regardless of the industry. I can easily find two examples from Start-up/Credit Risk (where +0.01 in AUROC translates into +1M USD) or Corp/E-Commerce (where -1% on MSE can easily translate into +3M USD only in a quarter) which illustrate well such impact.

如果您拥有这样的自动化的数据驱动业务，则无论行业如何，+ 1％的性能对模型性能的影响都是巨大的。 我可以轻松地从启动/信用风险(其中AUROC中的+0.01转换为+ 1M USD )或公司/电子商务(其中MSE -1％的收入仅可以在一个季度中轻松转换为+ 3M USD )中找到两个示例。很好地说明了这种影响。

Tomorrow…is weekend.

明天…是周末。

In the next, third and final part of this post, I will go through the reasons why this discussion exist on what are the true obstacles for a large-scale adoption of data science in industry. Stay Tuned for more!

在本文的下一部分，第三部分和最后一部分中，我将探讨为何存在这种讨论的原因，即在工业中大规模采用数据科学的真正障碍是什么。 敬请期待！

P.S.: I would like to personally thank to Fernando Costa, Sven Thies and Jihed Khiari the time they devoted on reviewing this post. Kudos to the three of them.

PS：我要感谢Fernando Costa ， Sven Thies和Jihed Khiari投入的宝贵时间。 对他们三个人表示敬意。

<< Zero-Stack Data Scientist — Part I, Beginnings

<<零堆栈数据科学家-第一部分，起点

Zero-Stack Data Scientist — Part II, The Rise >>

零栈数据科学家-第二部分，崛起>>