Introducing the AI + Design Series
A designer’s journey to discovery
Artificial intelligence can sometimes feel like an empty catchphrase, but the prevalence of discussion around AI points to its power and potential for transformation. AI is helping make advances in everything from science and medicine to art and fashion. At Zalando, it’s helping us solve many of our customer problems better than ever before.
There are several topics I want to talk about that warrant separate posts themselves. So that’s why I’m kicking off this Design and AI series with an introductory post about how I started learning about AI as a Product Designer.
Back in October I read about Google’s AlphaGo Zero taking just three days to master Go, the ancient Chinese board game.
It not only figured out thousands of years of human accumulated knowledge of the game — it invented brand new moves of its own that Go grandmasters are learning from today.
That’s when it hit me. People are learning from machines instead of the other way around! I had been aware of AI being used to imitate people (e.g. making art) and collaborate with people (by gaining a deeper understanding of audience and by creating new tools). But now with this Go example, I saw that AI has the potential to make us better than we could ever be before.
Joining Zalando’s personalization team
One of the first things I noticed when I joined Zalando’s personalization team was a photo metaphor of our team’s goal. The image showed a warehouse full of products transforming into a personalized boutique for each and every customer. Not personalized for customer segments, but actually for each customer.
I had a rough idea that machine learning could help us understand our customers better by uncovering insights through the use of data, but I wondered what else we could do. I also started learning about semantic web and ontology from my colleague Katariina Kari (we now play together in the Zalando string quartet). But wow, was there a lot to learn!
Diving into customer problems
Our user researchers are constantly gathering qualitative and quantitative insights, so I started to categorize some of the customer issues they were encountering. I decided to focus on three main problem areas that were relevant to personalization:
- Customers having issues combining separate pieces into outfits they could wear with confidence and that were in their price range
- Customers feeling overwhelmed with the amount of choice on Zalando
- Customers not relating to the products they’re seeing on Zalando
How might we…
So once we had identified the problems areas, it was time to identify the opportunities. I asked several teams for input on the following:
- How might we help people combine their old and new fashion pieces into outfits…and at scale for millions of customers?
- How might we surface the most relevant choices for each customer so the number of options isn’t so overwhelming?
- How might we be more transparent to customers why we’re showing them their specific recommendations?
Algorithms to the rescue
There are several teams currently focused on solving these customer issues. Teams who are envisioning the future of our algorithms to:
- understand fashion and outfits
- discern what each individual customer likes
- understand what a customer’s in-the-moment shopping intent might be
We want to inspire people, and perhaps even help our customers learn more about themselves and their style. This is a tough challenge, particularly because fashion is so subjective — style and taste are very personal things! Additionally, it’s difficult to predict a customer’s in-the-moment intent when a huge part of your customer’s fashion experience is happening elsewhere (both online and off).
A few of our teams created and released new outfits features based on past purchases and wishlist items, and we tried them out with human-curated outfits. The features fared pretty well in both user testing and A/B testing, so we moved on to testing out algorithms. We asked ourselves, how can we ensure an algorithmic outfit quality that is as high as “human quality?” And what is “human quality” anyway?
This was a new concern I had never encountered before as a designer, and it was exciting. Getting this right meant we were closer to what was impossible before — to personalize for each and every customer…and for millions of them!
That’s when the team decided to measure outfit quality by conducting our own variation of a Turing test with the fashion-savvy people at Zalando. We approached it this way because we wanted the algorithmic outfit quality to be as high as human quality, which because of its subjectivity can be difficult to define.
We asked ourselves, how can we ensure an algorithmic outfit quality that is as high as “human quality?” And what is “human quality” anyway?
This basically meant we’d throw both human-curated and algorithmic outfits into a friendly competition against each other, and I couldn’t help but think of my days playing the Street Fighter video game, me vs CPU!
If the algorithms’ outfits were able to match or outperform the human-curated outfits, that would be a good sign of being close to human quality. We were able to gain some insight into what categories and characteristics are harder/easier to combine, and we uncovered some seasonal trend considerations. Some examples were that it was algorithmically harder to create maternity outfits, and it was harder to combine some colorful and patterned shoes. I was most surprised with how I started to feel like the algorithms had their own personalities, communicated through their choice of outfits.
During lunch with a friend the other day, I had an “a-ha moment” while drawing on a napkin to explain what I do (the axes were inspired by personalization leads Volker Pilz and Saku Laitinen). I realized that the services we offer could be grouped into four quadrants:
- Automated + for individual
Auto-relevant for you
“All of these are tailored to your specific likes and in-the-moment goals.”
- Automated + for everyone
Auto-related to this
“If you like this, here’s more like it.”
- Manual + for everyone
Hand-picked for all
“Here’s what’s popular/in season.”
- Manual + for individual
Hand-picked for you
“A person will help you find exactly what you want.”
Some final thoughts:
- Quadrant 1 has the potential to solve some of our users’ problems better than ever before. This is where we crack how to offer personalized and even predictive content that’s unique to each person.
- Creating these scalable hyper-personalized solutions is not just about mathematics, but it’s also about how we inject our algorithms with a deep understanding of human values and needs to create products customers love to use.
- The data and tools we use to create is very human actually — they’re made up of user data, direct user feedback, and research insights from customers, stylists, and other partners.
This is only the very beginning of my AI learning journey, and there’s obviously a lot more to learn. For example, I’m currently thinking about different ways AI can be introduced into our product creation processes. Which ways will actually help us and why? Stay tuned for this topic and more in this “Design and AI” series!