Skip to content

Interview with Dr. Claudia Perlich, Chief Scientist at Dstillery


Dr. Claudia Perlich is a data scientist with twenty years of experience in big data analytics and machine learning. She is currently Chief Scientist at Dstillery, a marketing and advertising analytics company headquartered in New York City. Prior to 2010, Perlich spent over a decade conducting data analytics research at IBM’s Watson Research Center. She is the author of over 50 scientific articles on topics like data mining, social media analytics, and predictive modeling, and she teaches Data Mining for Business Intelligence at New York University’s Stern School of Business. Perlich holds a PhD in Information Systems from NYU, and an MA in Computer Science from Colorado University.

Interview Questions

[] You earned a master’s in computer science, a doctorate in information systems, and, as you said in an interview at the Women in Data Science Conference, you started working in the field before the job title “data scientist” existed in a formal sense. Did you consider yourself to be a programmer, an analyst, an IT specialist?

[Dr. Perlich] There were plenty of words for data science, but it didn’t have the public recognition it now has as something interesting, important, or sexy. The first time I ventured into the field of data was back in 1995. I just happened to stumble upon a course in artificial neural networks with no clue that it was about data. I wasn’t concerned with labeling what it was – I think labeling typically does more harm than good. I just really enjoyed what I was doing, which was learning about the world through data, and solving the puzzles that data presented.

Throughout that, I stuck to the narrowly, technically true definition of what I was doing, which was supervised modeling and, arguably, the next level above that would be machine learning. But, I didn’t really care as long as I enjoyed what I was doing.

It was clear when I was looking for a place to get my PhD that I wanted to continue in the vein of analyzing data and building predictive models. It was actually an unorthodox choice for someone with a background in computer science to go to a business school. Very few business schools at that time had a technical specialization. It was an accident of fate that NYU Stern School had incorporated machine learning, and one of my professors from ’95 had gotten a job in that department. That’s how I got there.

[] What was unique about the NYU program?

[Dr. Perlich] Information systems departments tended to be at the intersection of business, technology, and human behavior. And, in 1998, you didn’t find machine learning fitting into that. On the business side, the discussions were more about competitive advantage, IT investment, and how technologies like email were changing management interactions. That was the primary direction of information systems research. There were very few departments in the world where, in an information systems department, you actually found people who had specialized in machine learning and data mining for business applications. Stern was clearly different in that respect.

I wanted to continue to do machine learning. By the same token, I’m grateful that I got to do machine learning in the business school, because the focus was on applied problem solving, which suits me better than developing new fancy algorithms. Designing complex algorithms has its own merits, but my own personal interest was in finding interesting problems to solve more so than making algorithms faster or proving convergence, which was more of the focus in computer science programs that had machine learning.

What business school gave me was an appreciation for trying to solve real problems that were of true interest to businesses. And that set me off on the career path that I’ve been on.

[] Did you feel at the time that you also making an atypical career choice for a woman?

[Dr. Perlich] In all honesty, I don’t think that I have very often even asked myself the “as a woman” question. My internal notion of my identity was less defined by the fact that I am a woman than it was by my interests.

I’m kind of the classic tomboy. I played with the boys growing up. I spent my teenage years in weight rooms, lifting weights, which is very much male dominated. I went into my first computer science lecture and there were maybe ten out of 600 students who were women. So, I probably don’t have a lot of sensitivity on this issue.

In contrast, when I recently went to a Women in Machine Learning conference, it felt kind of odd to me to find myself in a room full of women. I’m not sure exactly why that is, but machine learning is so much a part of who I have always been that I have never asked myself if this is unusual for a women. I may have felt a little bit like the odd person out when I initially got into the field, but that may have just been about me being an introvert in a world dominated by extroverts. I had the sense that there was something that made me different, although today I think every teenager has that so it’s not that special. But, I never attached that feeling to being female. I just had odd interests and hobbies. And, for a time, after reunification, I was the East German living in West Germany. So, this whole notion of how you define your identity became much more internalized for me, and being a woman was not the most important part of that.

[] In your teaching at NYU, have you noticed any differences in the male/female composition of Data Mining for Business Intelligence classes over time?

[Dr. Perlich] I’m not very good at recalling what it looked like eight years ago because I don’t think I’ve always paid that much attention to it. The reality is that recently it’s been at least 30% female, and it has felt almost balanced to me. There are a few important caveats to that statement. This course, Data Mining for Business Intelligence, is taught in two versions. There is the managerial track, and there is the technical track. The technical track requires Python programming. My sense is that the technical track has a different composition of students for that reason. I may end up having a more balanced class because I am teaching the managerial track. The goal of the class I teach is learning how to manage data science, not to be a data scientist. So you don’t have to be able to program. The gender preference may be playing out more noticeably on the technical and programming side.

[] What kind of advice would you give to women who are interested in pursuing a career in data science and analytics?

[Dr. Perlich] There is the question of whether or not you should pursue a career in data science. And I am standing by the belief that there will be an almost unlimited demand for data scientists moving forward. I just can’t see any downside. So, if you have any inclination that this is the kind of work you would enjoy and that you would be good at, you should explore it.

What is gender specific about this? The first advice I would give is, don’t ponder the question too much. It really shouldn’t matter. At the end of the day, it’s what you’re good at that matters. And the more you ponder the “as a woman” question, the more tension you create for yourself around it. My experience has been that being a woman can be beneficial. It is possible that I was lucky, or just good at selecting environments where nobody seemed to care about me being a woman. I know there are other environments where there is more implicit bias in the culture. It’s hard to know which way the causality goes here: once you start wondering whether they don’t take you seriously as a woman, then you start having to prove yourself and it puts the question further into the foreground. It can be a bit of vicious circle.

Next point is that women may tend to have issues with is confidence. The short answer to that is you should be proud of what you achieve. Find out what gives you confidence, and whether there is a place where you can feel freer and more confident. In my experience, if you walk into a room with the certainty that you know what you are talking about, typically people don’t think beyond that. When you appear to lack confidence, people may start wondering. So, I try to mentor people to give them confidence in what they know and do. There are certain cultural biases that are attached to female behavior patterns and these may get in the way. I hate to say that you should behave like a man, because that’s not really the point. The point is to trust yourself and to figure out what helps you to feel confident with who you are and the work you do.

The other important thing about confidence is this. If you don’t have self-confidence, then you may perceive failure as a personal shortcoming. Whereas, failure in data science may just mean that the problem isn’t solvable. Maybe the data you have is just not up to the job. So, there are inherent limits and some things just don’t work, and it is important in data science to realize that that is not a personal failure. Working through what might be perceived as a failure, and appreciating failure as a learning experience or an insight is a valuable skill in data science.

[] That reminds me of a story I heard you tell at the Women in Data Science Conference about a machine learning application for breast cancer screening, and how it figured out how to choose patients who were more likely to develop cancer based on where they had their screening, rather than on physiological factors.

[Dr. Perlich] Yes, and this is some area where women may have an advantage in that there is good deal of intuition that’s helpful when you’re working with data. It’s not just knowing the limits of theories and statistics. It’s something entirely different. It’s a gut feeling you get when you look at the data and you get the sense that something might not be quite right.

I have often seen in computer scientists that there is a preference for black and white, for right and wrong. There is less appreciation for the gray in between. Data science, at the end of the day, is never right or wrong. It’s not like a sorting algorithm that’s either sorting or not. You build a model and it’s as good as it can be right now, which is somewhere on a spectrum. Having the intuition to navigate this not-quite-right/not-quite-wrong gray area is very useful. Most of the time, when I’ve found something wrong it is because initially it looks too good to be true. That is a quality I have seen in people coming into data science from fields other than computer science and mathematics. Often, physicists or people coming from biological sciences have this quality. And, I have seen many women coming from these fields into data science and having that intuition.

[] You worked at IBM, and now you’ve moved to a smaller company where you’re doing a different kind of data science. How would you describe your current work?

[Dr. Perlich] A very good question is, why do you leave IBM Watson to go into advertising technology. What I find fascinating about advertising, is it is the perfect playground or sandbox for data science. From a data perspective, you get a huge amount of a large variety of data in advertising. It’s possibly the closest thing to what is hyped as “big data” that you can find consistently.

With the technology under the hood, it’s an advantage if you get to build something from scratch. Even IBM struggles with legacy data systems that they can’t replace. Once you have data technologies that are twenty years old, which is also true in areas like banking and many other places, it’s hard to move on to the new tools that are out there. We started about eight years ago with big data technologies that allow us to build very effective data infrastructures that you won’t see in many other industries.

Also, we have the ability to run experiments, to test, and to learn very quickly. The turnaround time if you are trying a new algorithm, for example, is quite short. In advertising you have the law of large numbers: there are so many ads being shown all day long that you can quickly learn something. And, truth be told, we’re not talking about breast cancer here. You can be freer about your choices because failure in advertising is not the end of the world. I see it as a great place to research how algorithms behave, and what kinds of biases they carry – all of the questions we are now asking in important fields like predictive policing and hiring practices. In advertising, you get a good glimpse of how AI really works in different applications. So, this is a really fun place for a data scientist to be at this point.

[] So, are you’re building models for advertising campaigns, or for another aspect of advertising?

[Dr. Perlich] Without getting too technical, as people use their digital devices, they may or may not be aware of how often they are being exposed to ads. The majority of the ads you see or don’t see are being auctioned off in real time. Every day we see 100 billion auctions happening where we can bid on showing people ads. That means making 100 billion decisions a day. There is no way to do this without machine learning. So, you have to act very quickly, and you are able to hyper personalize the ad placements. Given everything we know about a person through the data we have, how can we determine if that person is in the market for a new credit card, a Maserati, or new Microsoft software? This is an interesting problem for data science and machine learning. The big questions are: can we predict what people are actually interested in?; and can we predict what people are likely to do in the near future?

What we’ve done at Dstillery is build an infrastructure that anonymizes information, while retaining the data that may determine who would be a good candidate for a specific ad. We have the technical infrastructure to cope with huge amounts of data, and the machine learning models for all kinds of purposes. So, at IBM in that breast cancer screening model we mentioned earlier, that was six people spending half of our working hours over three months to build the one single best model. Today, I’m operating in a much smaller company where a team of 13 data scientists are building ten thousand models every day. That’s how far we have come in just eight years.

[] It’s clear that you need to have a deep understanding of the technology you’re using in data science, and you’ve pointed out that instinct or intuition is also a valuable skill when evaluating the results. It would also seem important to have the kind of experience that would lead you to ask the right kinds of questions. Is that accurate?

[Dr. Perlich] I would like to say that you are right about that, because that should be true. In advertising, the question is often dictated by the customer, whether or not I believe it is the right optimization criteria. So, educating marketers on what the right questions should be, and which of those questions we can and cannot answer effectively is something we have to think about. I think we have made some progress in that area.

Intuition can be very valuable when you are trying to debug something. Let’s say you’re looking at a particular campaign and it is not working. You need to consider out of the 100 different variables that went into it, which is most likely to cause the problem that you are seeing. That requires a deep understand of the full data pipeline, from the early data collection process, up through the execution and the decision logic.

Also, we build models all day long, and unless something is broken the thing does exactly what it is supposed to do. But, very often we are looking at the question of what else we can do with this data we have. There are interesting thoughts about new products, insights, and solutions we can come up with based on the data we have. Can we come up with causal claims for attribution? Can I prove that the campaign drove people into the car sales lot? My sense is, probably not, given the data quality we have right now. Can I predict who will go there? Yes, I can, because we have the data to support that. This is another area where having some intuition about what data can do is an advantage.

[] Since we’re talking about data and causal factors, do you feel that the field of data science has evolved in order to attract more women, or could it be that the culture in is evolving simply because more women are moving in to the field?

[Dr. Perlich] My own experience has been that since my PhD I have ended up in environments where the team around me was closer to balanced than most places. My PhD advisors had three data scientists graduating under him when I was there, and we were all women. At IBM, half of the predictive modeling group, or maybe 30% were women. And, now at Dstillery, we have at least 30% women in the data science team. I’m not sure why that is, or if that is a representative example.

Interesting fact: it seems that statistics has generally had a higher number of women in it than computer science. I think that there is something of this sort going on, and to some extent it is a beneficial side effect of the tremendous demand for data scientists. If you are having a hard time finding data scientists to hire, you’re going to hire someone who is qualified. It won’t matter if that person is a woman. By the same token, if you are a women with technical interests, there’s no questions that data science is one the most promising areas to be. This is a field in which you can’t be picky about who has talent. That may push whatever gender biases people have into the background. Talent is rare, and you are glad for any that you can find.

Also, I think women tend to be very good communicators. Data science is one of these areas where they talk about the “unicorns,” or people who are supposed to be able to do statistics and math and present or communicate very well. Communication into the business side can be a real challenge in data science. Being able to do that well may not fit the idea of the hardcore data scientist working on math equations. But data communications is a challenge that involves being able grasp the bigger picture, define the problem, and explain that well. That is an aspect of data science that has been welcomed and recognized as an important component.

[] Thank you for taking the time to do this. Do you have anything to add?

[Dr. Perlich] I appreciate the fact that your questions went beyond the more popular ideas about women in data science. I’ve struggled with what role I want to play in this discussion. If I can be a role model of sorts and help younger women find their confidence, I want to do that. At the same time, I think people should just do what they are good at and what they like to do. For me, the goal is not that we get to a 50/50 world in data science with women and men, but it would be great to get to a point where we don’t have to discuss it anymore. That should be the goal, and that would make me happy.

[] That is the paradox: in order to get to the point where we don’t have to talk about this anymore, we sort of have to talk it about a lot more.

[Dr. Perlich] I know. But it’s important to keep in mind the ultimate goal.