Building data science products in multi disciplinary teams

For the last three years, I have been working on different data science projects at Zalando, helping our more than 24 million customers find the most relevant items in the assortment we have. Along the way, I have learned how to scale data science, or how to build a new personalization product from scratch. Thanks to my experiences, I am a firm believer in having dedicated and autonomous multi-functional teams to solve complex problems, especially when they involve learning.

As data scientists, we are used to looking at problems from a data perspective, which has helped the teams I have worked with gain huge amounts of domain knowledge. We strive also to make data-driven decisions, where running A/B tests or doing online and offline evaluations of the models we build are some of the most important tools we have. However, what does it look like to work in a multifunctional team?

In a Zalando team, we usually have one or two data scientists, one or two engineers, a product manager, a designer, and sometimes a business developer. The details change from team to team, but you get the picture. Not all of these people are dedicated to the team 100% of their time, sometimes a designer can work with two or three teams, depending on their areas of interest. The main advantage of working with this setup is that we are able to tackle uncertainties and risks from many more angles, and way faster than on a researchers-only team.

Something I have learned when working with designers, is the many advantages of early testing and prototyping, and their customer-centric approach to problem solving. Moreover, because they tend to work in different products from similar areas, the knowledge transfer usually happens more naturally and also faster, and completely changes the way we work. When working closely with our copywriting team, we learn how to communicate our products in the right way for our customers, and working with engineers we learn how to make sure to build machine learning solutions that scale; ones we are able to operate.

A very good example I have previously written about is the latest product I was in charge of building, where we were able to collaboratively design a prototype to solve our customer problem of, “How can we make recommended content more transparent and relevant to our customers?” We did this in four days, writing only a minimum amount of code. We built six personalized prototypes for user testing, by manually adding “recommended” content into a static version of the Zalando App. Instead of using an algorithm, we “faked” the algorithmic result by using human expert curators to choose which content would be shown to each customer.

By faking the personalization part, we were not only able to understand our customers expectations about our product, but we also saved months of development of an algorithmic solution that was not what the customer expected. In particular, the feedback we got from our customers was far more specific and natural than when using non-personalized prototypes. For example, instead of asking someone “imagine you love leather jackets and we recommend you matching boots,” we can know beforehand that they bought a leather jacket last week, and we created “recommendations” of the boots we thought would better match their style.

Working in this environment is also aligned with our autonomous teams. During the process, everyone involved gained customer understanding and domain knowledge from the problem we are trying to solve, something extremely valuable for data scientists. Moreover, iterating on this is way cheaper and faster than iterating on A/B test cycles, even when we have a really strong testing-as-a-service infrastructure.

This is only one example that shows how much I like working with people from different backgrounds and functions, which also proves how important diversity is for building great machine learning products, especially in a B2C market that operates on a European scale like Zalando does.

* Humberto Corona is a product specialist and data scientist in Zalando's Fashion Insights Center in Dublin. A regular contributor to the tech blog, Humberto is a finalist in this year's DatSci awards, where this piece was originally published. Ana Peleteiro Ramallo took the Data Scientist of the Year title in her role at Zalando in 2017.*

We're hiring! Do you like working in an ever evolving organization such as Zalando? Consider joining our teams as a Machine Learning Engineer!