Each day at Zulily we add 9,000 products to our online store and process more than 5 billion clicks from online shoppers. That is more virtual inventory than you’ll find in the warehouses of many retailers, and it’s by design. We’ve built a supply chain where we hold only some goods: most of the time, we don’t purchase inventory until our customers have, so we are able to pass down savings from our unique supply chain down to our customers around the world. To the customer, that means a constantly changing and new shopping experience. Our system works only if we can ensure that both shoppers and suppliers move quickly. To do this, we’re cloud-based, data-driven, customer-first, and obsessed with machine learning—pretty cutting edge.
The way we get to all that innovation, however, is by staying close to a few basic rules that we amplify with cutting-edge technology, particularly when it comes to the way we look at the utility and purpose of data. And though Zulily may specialize in ecommerce, every technology professional interested in AI and ML can benefit. Here is a short list:
1. How well you describe your data is key
Obviously, we have a lot of data, and we believe strongly in data analytics and ML to power our business. A critical part of this analysis is the tags we put on data, and how well we describe things with those tags.
For example, we sell a lot of fashion. Frequently, we’ve found that things that matter to consumers, like the length of sleeves on a shirt, or the height of a shoe’s heel, are not inside standard product descriptions. Typically, this hasn’t been an aspect of clothing descriptions online—suppliers may only provide that level of detail 20% of the time or less. We could manually input information by having someone examine images of each product before it goes live, but this would be time-consuming and inaccurate.
Using AutoML Vision, which we use to train our own custom ML models from our image data, we get much closer to expressing data about our products that our customers want to make a buying decision. We train custom ML models to classify product images according to labels we define, such as sleeve length. This enables us to include sleeve data in our product descriptions more than 80% of the time.
This isn’t a problem that is unique to retailers. Take a company like real estate company Zillow, where I previously worked. Companies like them have a slower-moving inventory and work inside a different business model, but face similar issues around turning photographic information into floor plans. You also see leaders in construction-related fields use AI and ML to determine the types and amounts of stone in a quarry, based on drone and mobile device video footage.
2. The same data can serve both customer needs and business needs
Returning to the sleeves example, we wanted that information because it mattered in customer choice. So sleeve length is data labeled in a way that matters to a customer need. In addition, Zulily offers discounts of up to 70% on our products, with limited availability. This means that our products aren’t around for long, so we want customers choosing swiftly, and feeling confident in that choice. So you could say discounts and stock levels is data that also serves a business need.
I’d challenge anyone working in data to go through this exercise: Look at how it’s organized and think about how directly it is serving a customer or business outcome.
This can take different forms, meaning the same data will serve different ends. On the one hand, Zulily’s challenge is volume and rate of change, and much of our data needs to address scale, in the form of new selections for customers based on interests, or managing diverse and changing supply chains in a flexible way. On the other hand, we need to move goods quickly, and want customers to understand that items in their shopping carts can disappear.
Even as we’re bringing items into stock, we’re measuring and predicting how fast things run out, so we need data that serves both business needs. Away from its ML uses, Google Cloud helps here too, since it’s easy to scale up and down compute and storage, orchestrate that programmatically, and manage and manipulate our large datasets.
Keeping your data in shape to meet your business needs isn’t just good for operations. It helps get your data teams thinking in terms of the business, and your software developers cognizant of how the data ought to look, which leads me to my final point.
3. AI is software and science. Successful projects use both.
For the time being at least, successfully managing corporate AI means navigating between the cultures of data science and software engineering. There is a big difference: software engineering has far more certainty; it is about predictability and deliverables, according to a timeline. Data science is about testing ideas, unpredictability, and the necessary failures that accompany most learning. These two can end up in conflict.
This is a particular challenge in retail, which has tight margins, seasonal comparisons, and fiscal restraint. Everyone wants growth, but they don’t want to risk too much in their experiments. Sometimes, to gain insights, it’s necessary to lose 1% for the sake of experimenting—necessary for embracing the future.
This is a razor’s edge we, and you, will have to walk. The solution is to minimize the impact of experiments by designing them to learn as much as possible as quickly as possible, either by impacting a large portion of the experience for a short time, or a small proportion of the experiences for a longer time, as the significance of change to the customer experience or system behavior dictates.
In addition to that of science and software engineering, it is worth noting that Zulily must also navigate another “two cultures” dynamic: that of machine-generated content and human-generated content. Both have their strengths. In general, people are better at generating original descriptions of products, but natural language processing can quickly produce brief summaries from that. ML can forecast and trend, but typically buyers want to draw on their own intuition as well. The trick is in finding the right balance—we tend to lean towards the machine while giving our buyers—who are experts in determining what customers want and interpolating that with styles from brands and vendors that are coming—some input as well.
Remember: Technology works best when it serves a human need
As I said at the top, we are cutting edge, with a cutting edge cloud. But there’s something even more important than technology itself that we remember: Technology works best when it serves a human need. Customers need to understand a lot about a product if they’re going to choose it quickly, for example, but it would take way too long for Zulily to manually annotate each image. Worthwhile tests—like using custom ML models to classify product images—can help us get much closer to what the customer wants, but they have to be executed in a way that doesn’t potentially affect customer experiences.
We believe in AI, ML, cloud computing, mobility, new delivery mechanisms—all of the cutting-edge tools. We believe there’s lots more to come. But we’re succeeding because we’re concentrating on the essential truths of retail—serving our customers with purpose, and as individuals—and that won’t change in the digital era.
Keep learning: Learn how Google is supporting retailers across the globe during COVID-19.
About the author
Olly Downs, VP, MarTech, Data & Machine Learning, Zulily
Olly Downs is a machine learning scientist, seasoned technology leader and serial technology entrepreneur, credited with bringing advanced analytics and machine learning methods to bear as the creative spark behind numerous early-stage technology companies. Teams Olly has led have built business-enabling technologies that have returned almost $2 billion in business value on $250 million in venture capital investment. Learn more about Olly here.