Personalization

How THE YES Approaches Personalization

The choices that consumers have in almost every product category are greater today than ever before, but this abundance can make it more difficult for them to sift through options to find what they really want. Julie explains why personalization helps to bridge that gap, as well as the difficulties of executing it properly.

I think what we've seen with the growth of the internet is there's just so much out there that really the next generation of the internet is about how you curate for each human what is most relevant to them and how you reduce the overwhelm.

I think that's a problem of the internet at large. We believe that, in commerce and fashion specifically, our focus on this category of fashion gives us an advantage. While there are these big aggregators out there trying to pull together recommendations for you across tons of categories, my belief is you need to be deeply entrenched in the vertical you're trying to personalize to understand what matters in the category and what matters to the consumer.

If you look at a Pandora or a Spotify and all the time and years they spent understanding music—they too had humans who were first creating the dimensions of a piece of music and then using machine learning to identify these patterns and match them in other pieces of music—that's what we're doing for fashion.

It's an enormous category, the business of fashion. Any company that is trying to do this really well takes real TLC, dedication to the vertical, to understand it, to know it very well to build algorithms that can truly give you recommendations that make sense.

My general view and hypothesis is that, if you're going to build a really strong system around personalization, you need to have the largest assortment on Earth. You really need to have enormous choice in order to make the recommendation great.

On the flip side, if you're building the world's largest fashion store, you need to provide the customer some way to make that experience not completely overwhelming. I see the intersection of a very large-scale business with very intelligent machine learning and algorithms to make great recommendations, to be that sweet spot.

We had to build technology to allow us to scale and work with hundreds, and ultimately thousands of brands, including their entire catalog. The more brands and the more products we get, the more data points that we have, and the better our recommendations get.

There's human judgment involved at the foundational levels. We used a human or a group of humans to build out the taxonomy that originally defined the labels for each product. You need someone who's really an expert in the category to know.

We also use fashion expertise to build what we call collections for trends. A human will determine what the trends are, and then the machine will sort of seek out all the products that match that trend. One woman and another woman might both get camouflage as a trend, but they will get different products based on brands and price points and other factors that matter to them for their personal interests and tastes.

Having the right experts on hand can keep THE YES up to date with fashion brands and trends. This knowledge base, while big, is relatively stable.

However, consumer preferences are individualized, subjective, and constantly changing and therefore present a more complex problem.

We will learn more about how THE YES learns about its consumers’ preferences next.

Learning About Consumer Preferences

Julie discusses THE YES’s user interface. After completing a short questionnaire to input preferences, users scroll through clothing items and indicate what they like or dislike by pressing “yes” or “no.”

The app only covers women’s fashion, but does so in great depth. It has a sophisticated and personalized recommendation engine that continuously improves with use. THE YES gamifies the shopping experience so that looking for clothes is as valuable and fun for customers as buying them. 

If you ask most women when they're shopping online, they have five tabs open for different stores with different assortments in each store based on what, say, Nordstrom has bought versus what Neiman's has bought from each of the designers. You can often find yourself twenty pages deep into a search for a dress for some event, as an example.

What we're trying to do is really surface the things that are relevant for you. We have the entire collection from every brand we carry. The yes and no is kind of our love language. It's the way that the customer expresses whether she likes something or she doesn't. It's our branded way of thumbs up and thumbs down.

But it's really helpful, and it provides a couple of layers of signaling for us. It's really the interaction that the customer has with the product. It's a way for her to save all the things she's interested in. It's a way for us to learn what she likes and to learn what she doesn't. No is as powerful as yes.

It's actually a really satisfying experience as a customer to be able to remove the things that you don't like. It's funny, my daughter, who's eighteen, every time she says she goes to the website she likes, she wishes now that she can yes and no the product because she wants to remove the stuff that's not interesting and relevant to her.

In situations where understanding consumer preferences and behavior is very complex, companies have found success in using various artificial intelligence (AI) techniques, which allow for an effective analysis of large amounts of complex data.

One subfield of AI that has been particularly useful is called computer vision. Computer vision deals with processing digital images and videos to extract high-level understanding from them.

As Julie explains, the nature of THE YES’s work necessitates the use of this technology.

A lot of people can take an image and use computer vision to show you other images that might look like the product. So Google and Pinterest have done really great work on the computer vision side. But if you think about clothes, it's much more than just what does it look like. It's what is it made of, how does it fit me, what's the brand, what's the price.

Our algorithm factors in style, as well as all of these other factors that are really important when you're buying an item of clothing or an accessory. When we think about style, we think about it in layers. So there's kind of first, the family of style that something belongs to. So is it boho, is it classic, is it edgy? We ask sort of one layer as you're onboarding around general groups of styles that you prefer. Nobody is all one thing; everybody's a mix.

That's the first layer. And then the second layer is, we have actually broken down every garment into all of its style attributes, so we can start to learn patterns. Are you someone who doesn't like v-neck or doesn't like crewneck or doesn't like boatneck? One of the best questions we ask is this polarizing screen where we say, what would you never wear. For me, I'm a fifty-year-old woman, I would never wear a crop top.

These are things where upfront, I can give you more polarizing styles that give you a very clear signal around what not to show me. Move that out of my way. If I searched specifically for it, for example, if I wanted to buy a gift, I could do so, but it's never going to pop up as a recommendation for me.

We've sort of broken products down into all of their dimensions. Then we've rolled things back up. And as you're saying yes and no, we're starting to learn signals. If you first have answered the polarizing style screen, we'll know never to show these things. Then there's lots of nuances.

When it comes to color, for example, you may not generally like a color, but seasons change. If something becomes really popular, you may decide you want that. So color is not as absolute, for example, and we've learned that. Even if people say they rarely wear colorful things, they still sometimes might want to see it.

What we do is, we have algorithms that help us with diversity. We can show you the things we know you'll like most up top, but we'll sprinkle in some things that may interest you, whether it's an unusual color or it's a style that's maybe new and we don't have a read on how you like that.

We've really deconstructed all of the categories of fashion into hundreds of dimensions that come together to form our taxonomy. Then we're learning patterns over time. We're also learning relationships between things that might be a cold shoulder style and an off-shoulder style, which may be more closely correlated than a turtleneck and a cold shoulder style. So we also learn correlations of style.

Personalization in any category involves understanding the products, but the other aspect of personalization is the preference of the consumers themselves—another thing that can be difficult to pin down, as you will learn from Julie. 

Julie explains what steps THE YES takes to stay ahead of the evolving preferences of its consumer base.

We think about style as always evolving. So we definitely have ways to enable the customer to explore new styles as she's shopping. Our home feed, or our home page on the website, are really the areas to surface new ideas and get your reaction to it.

It's not that we're always trying to find the exact match to what you wore last season. We know that that's not how fashion works. It's a combination of the fact that the product is always changing, so just by nature, we're always surfacing new products for you, and the fact that we know that trends evolve.

We have sections that are products similar to things you've liked in the past, but we also have sections that are new trends that, we think, might be interesting to you. We take signals from past experiences and other people to be able to surface the right products in kind of new trends.

We're always testing new styles that we think would be relevant for you, alongside the tried-and-true styles that we know you might like.

Algorithmic Personalization

For THE YES’s platform to work well, it needs to incorporate data from a large variety of customers as well as from a large variety of brands.

Julie will now explain which of these inputs THE YES chooses to prioritize in its algorithm, and why.

Our initial hypothesis has been getting more brands is more important because we can understand enough about the product set and we can understand by asking each user very high-intent, high-signal questions what matters to them. And the more product we have to choose from, the better the match can be.

Over time, as we get more users, we'll benefit from the advantage of collaborative filtering. Our initial algorithms were built off of the consumer's input about herself rather than understanding other users. We think that will be additive, but we needed to start with a bigger catalog first.

As Julie observed, data on consumers' fashion preferences are critical for the algorithm to offer personalized recommendations to them. This requires rich data from as many brands as possible to learn about a consumer's preference. THE YES has therefore decided to prioritize acquiring a vast portfolio of brands before acquiring a large number of customers.

Now, the question is, how can the company best work with these large volumes of brand and consumer data? There are few important technologies that THE YES, or any brand wishing to do personalization, can draw upon. New artificial intelligence tools, like computer vision and natural language processing, can be useful in filtering and sorting this immense amount of data.

Julie mentioned computer vision in an earlier video and pointed out that THE YES required a more complex algorithm than computer vision alone could provide. Computer vision uses visual cues to identify and analyze digital images. For instance, a computer vision program could look through the online catalog of a fashion designer and group products by certain attributes, like color, clothing category, and cut.

But as Julie explained, this algorithm must be augmented by other factors that are relevant to its consumers, such as price, material, and subjective tastes. The team at THE YES decided to integrate another AI technology—natural language processing, or NLP—which attempts to understand text in a way similar to humans. THE YES's algorithm used NLP to aggregate and learn from words that users enter into the app's Search function.

After getting a large portfolio of brands on board, THE YES started acquiring customers that provided a new source of data to help personalization through collaborative filtering, a technique that Julie mentioned. Collaborative filtering uses the overlap in consumers' preferences, or collaboration, to predict their preference for new items. For instance, if both customers A and B like the same nine skirts, collaborative filtering would reason that customer A would also like the tenth item chosen by customer B.

It is important to note that no direct knowledge of the product is necessary for collaborative filtering as all the filtering is done by matching customer preferences. A different approach, called content-based filtering, uses categorical tags or features of the product to create recommendations. For example, if a customer expressed interest in several pairs of red shoes, content-based filtering would present you with more products that match that description.

Julie mentioned this content-based approach earlier when she noted that the company's algorithm uses different layers to learn about consumers' preferences. For example, according to Julie, she would never wear a crop top or off-the-shoulder dresses.

Here’s a summary of the techniques introduced:

• Natural language processing: an AI technique used to analyze and understand text and language in a way similar to humans
• Collaborative filtering: a technique that makes use of consumers’ overlap in preferences to predict their preference for new items
• Content-based filtering: a technique that uses categorical tags or product features to create recommendations for consumers

Content-based filtering has some important strengths. It doesn’t require data from other users, so it can be easier to start out with than collaborative filtering. It can capture the specific preferences of a user and can recommend niche products that few other users prefer.

Content-based filtering requires a high level of domain expertise and a large amount of labor early on to apply many useful tags to products. While some aspects of clothing are obvious and objective like color and size, others can be much more difficult to describe in ways that translate between users. Descriptions of style can be very subjective. In addition, the model can only recommend based on existing users’ preferences.

Collaborative filtering has some strengths as well. First, it can be effective in helping consumers become aware of new products, as the recommendation is based on what similar consumers purchased. Collaborative filtering is also appealing as it does not require extensive domain expertise or product data, as the recommendation is simply based on what others purchased.

However, collaborative filtering is limited by your customer base, as a smaller customer base creates a smaller sample size from which to draw correlations of customer preferences. It is therefore less accurate at the early stage of a startup when the company has a limited number of customers.

Collaborative filtering can also lead to a lack of product diversity because it has difficulty recommending completely novel products with no user data. This feature can be especially limiting in creative industries such as entertainment and fashion.

Collaborative and content-based filtering use different methods to filter, categorize, and personalize products, but they are not mutually exclusive. In fact, companies like THE YES often use both approaches because they complement each other’s strengths.

Personalization and Engagement

THE YES’s wide and varied catalog–and its algorithm that allows customers to navigate it–work together to provide an experience that THE YES hopes will prove valuable beyond those moments when customers are actively looking to purchase a product.

Here is Julie on how THE YES’s efforts towards personalization led to more engagement and higher retention rates.

If our business works, then we have a materially higher retention rate because we've built a system where the investment a customer has made is something that brings them back. And so, in theory, we're still playing this out, but we've seen this in the early numbers. Customers come back more often than they do to other stores they shop online because they're interested in looking at the recommendations, even if they're not actively shopping.

So there's an engagement piece. The feeling that we all have when we feel heard and understood is a pretty powerful feeling. To the extent that we can be your one-stop shop, our goal is to make the entire experience so much better that you really are retaining people because they find what they're looking for, the service is great, and they feel like they're going to find what they love more quickly—and be able to keep track of everything they own, pair things together.

We have a lot of thoughts around how to make this solve multiple need states for the shopper. And it allows us to continue to grow and scale and acquire customers cost-effectively in part because we have enough of an assortment that we can also retain customers. And we build a differentiated enough and strong enough of a system that it's a great experience and customers have a reason for wanting to come back.

The more time they spend on our platform, it's an investment that pays off. And so it gets better and better, and the experience gets better.

Personalization tends to be most useful when your product or service is difficult to describe and the variety is large. Fashion tastes are subjective, difficult to describe, have a lot of variety, and are always changing. This makes personalization especially important for THE YES. The more objective and quantifiable product differences exist in a category, the less personalization can add to customer experience.

When shopping for a computer or a car, for example, it is easy for customers to decide for themselves exactly what they would like. Algorithms provide limited value in such scenarios. Personalization can offer more value in categories where product difference is more subjective or abstract, like wine, fashion, beauty, restaurants, or movies. Personalization can be a helpful tool in aligning product distinctions with consumer preferences.

L'Oreal, for instance, offers a service called ModiFace, that provides personalized digital service to L'Oreal consumers. As we learned before, ModiFace uses augmented reality to allow consumers to virtually try on products or diagnose skin concerns by using its app on a smartphone. Users are able to customize their style or skincare choices to their personal complexions.

Stitch Fix also relies heavily on algorithms to understand consumers' preferences in fashion. In fact, its website includes an excellent algorithms tour, which explains how its algorithm works. The tour describes how Stitch Fix uses content-based and collaborative filtering to personalize fashion choices for consumers. It also works as an engaging piece of owned media marketing for the Stitch Fix brand.

Amazon launched a service called Prime Try Before You Buy, that allows consumers to choose up to six items from clothing, shoes, jewelry, and accessories, and try them free for seven days. Consumers can return the items they don't like, and are charged only for the items they keep. Amazon algorithm continuously learns from the data of what these consumers bought, and what they returned, to improve its prediction of consumer preferences.

In addition to the beauty and fashion categories, personalization has become essential in the online streaming industry. Netflix uses a combination of content-based and collaborative filtering to personalize a user's experience of scrolling through its content library.

However, personalization is not always easy to implement. It requires a large amount of customer data, and not all customers are willing to share their personal information with companies, even if it is used to provide them with a more seamless experience. Misuse of customer data by some companies has eroded their trust in this area.

To conclude our discussion of personalization, let’s turn back to Julie, who will remind us how personalization fits into the larger context of e-commerce and digital marketing.

I do believe that AI and personalization will be one of the cornerstones of the next decade's evolution with consumer products. I think it's going to have to be done sort of industry by industry, understanding what the needs are and the different need states are of the customer as they're shopping for whatever that category is. But I do believe that this idea of e-commerce getting smarter and learning from the actions that you're taking will become the foundation of all e-commerce and shopping over the next decade.

I remember when Dan Nordstrom hired me in 1999, and he said to me, it's all about, e-commerce is all about personalization. And at that moment, we didn't even have products selling on the website. So we had many more steps to get to. But in theory, you have this layer of technology between you, the retailer or the seller, and so there's no reason not to be able to learn from it and use it and make each user's experience more relevant to them over time.

It is the promise of technology, and I really believe we're sort of at this tipping point where that can now be done at scale in a much more intelligent way.

Whether through the tech-driven personalization approach or through some other tactic, marketers often conceive of customer engagement as a one-way street, with companies reaching out to their consumers in order to reduce churn.

As you will learn next, however, engagement can go both ways. In fact, marketers can reap great rewards not just by telling stories to their consumers, but by getting consumers to share their own stories.