Learning Personalization on TikTok

New data science helped them build an addictive user experience

One of my favorite podcasters was dismissive of a colleague’s suggestion to try TikTok. The 40-something guy said he’s too old and doesn’t have the patience to look at silly short videos. On a Friday afternoon he eventually relented and committed to giving it a try. On Monday morning he arrived in the studio looking haggard from having watched videos non-stop all weekend, and declared himself a convert. He described the AI-powered TikTok experience as mobile phone crack cocaine.

Calling TikTok’s US growth explosive is an understatement. In less than two years it has grown to over 100 Million US users who, amazingly, average about 55 minutes per day of usage. What’s its secret? A new genre of data science referred to as an interest graph. Let’s take a look at what they’re doing and examine how it can be applied to retail personalization. Spoiler alert: it’s particularly well suited for use by apparel and footwear brands.

Yesterday’s Data Science: The Social Graph

Let’s start in the pre-interest graph era, aka present day. Most all personalization systems are based on concept called collaborative filtering. This is a technology designed to discover data intersections. On a web site it generally appears as ‘people who clicked on Product A also clicked on Product Z.’ The Facebook social graph is also a form of collaborative filtering, essentially saying Jane and Sally are connected, and they also seem to share a connection to Tammy. Collaborative filtering has been the state-of-the art since the early 2000’s and is a core technology used by Facebook, Netflix, and Spotify. In fact, Netflix offered a $1 Million prize for anyone who could deliver something better than collaborative filtering. Over 40,000 teams from across the globe, many made up of talented amateur hackers, largely failed to beat the collaborative filtering standard.

So what’s changed? A key thesis of the social graph is that you’ll like the same things as your friends. In essence Facebook is giving you personalized recommendations from your friends. If you ‘like’ a movie this like will be shared with your friends.  Where does it come up short? Let’s say your passion is day trading and your friend’s passion is documentary films. It’s unlikely there will be many intersections. The social graph largely aggregates people demographically given that your friends share demographic attributes with you. Facebook’s business model is based on earning revenues on clicks, so the theory is you’ll click on things liked by your friends.

The Next Big Thing:

An interest graph is different. It links people together who share common interests. It’s not social, so the platform doesn’t bother letting you know who and why you’re clustered together. In TikTok’s case it begins showing you videos that are watched by people who share your interests. What’s an interest? It’s not entirely clear how their machine learning system defines an interest, although its users say the recommendations are addictive. Somehow it creates an abstract interest out of the 60 second videos of your friend lip-syncing a hip-hop song, making the machine a little smarter about recommending the next video.

How does this relate to retail?

At PreciseTarget, after analyzing over 5 billion consumer purchases, we’ve learned people have notably weak affinities to product brands. Consumers liberally drift between product brands, buying Nike this year, and Adidas the next. But, we’ve noticed that consumers are remarkably consistent in buying products that are similar to each other. Let’s say you have five pairs of jeans in your closet. They’re not all from the same brand, but they share common attributes. Their prices are similar as are the styles, colors, and fit. While you might describe the jeans matching your personal taste, our machine classifies it as a particular interest in jeans. Each day the machine looks at the entire universe of jeans, including the newest models, and ‘fits’ them to consumer interests.

How do we define an interest? Much like TikTok we don’t publish the recipe to our secret sauce, although clearly the attributes of the products you buy are instrumental in defining your interests. When our system machine learns a purchase, it’s looking at the attributes of the purchased product. For apparel this is typically 25 to 30 attributes, including price, style, size, color, fabric, cleaning instructions, country of origin, and even minute attributes like color of its buttons. Our comprehensive testing has proven that if you target customers by their interests (tastes) it will far outperform conventional brand repurchase models. Historically, we have delivered 25% higher click through rates in A/B testing against conventional targeting data.

Recommendations for the Future

We’re encouraging our retail clients to move closer to their customers by understanding their interests. We’ve started providing interest data to customers as an overlay to their CRMs, significantly augmenting their customer insights. We’ve also packaged the data in the form of targeting audiences for use targeting customers for new acquisition. This enables our customers to leverage our vast reservoir of knowledge about consumer interests, as well as our advanced machine learning and data science. We’re delivering advanced data science in an off-the-shelf form, enabling brands to more rapidly transition to a higher online mix.