Personalizing Fashion at Scale, Spotify Style.
While most consumer tech companies talk a great deal about Personalization, as a valuable customer engagement strategy, very few have made meaningful strides in delivering on that promise while generating serious business value.
Companies that stand-out in this crowded race seem to have a few things in common — strong digital DNA, data at the heart of all they do, and a large appetite for failing fast and learning.
With my recently acquired obsession to crack the “personalization code”, I started asking whether we could borrow these principles, of data science and design, from the pioneers and re-imagine personalized shopping. And could that deliver an addictively delightful customer experience?
In my quest to answer this question, I compared my experience across my most frequently used platforms — Amazon, Netflix, Google Now and Spotify. While all of them did a great job of improving my experience over time as they learnt more about me, Spotify was one that really stood out.
If you love music, then there is a great chance you know what Spotify is. For everyone else, Spotify is the leading music streaming service with over 140 Million active users, of which 60 million are paid as of July 2017. It generated $2.2B in sales last year and was valued at $8.0B in the last round of financing. But these impressive numbers are not what makes it special. Despite the fact that similar music services are being offered by the behemoths like Apple, Google, and Amazon, it still enjoys a special customer bond and loyalty.
So what makes Spotify special?
Its secret sauce is the algorithms powering the underlying software that brings you a personalized playlist of music. For those of you who are old enough to remember the era of music tapes, Spotify is the equivalent of a tape recorder that has the ability to create limitless “mix tapes” of your favorite music, without you having to put in the hard work.
Mixtapes were special since you hand-picked each track, created your own label — for example: “My love-songs hit list”, and could even gift it to your special one as a gesture of love. It was one-of-a-kind, no record company sold such a tape, was extremely personal, and hence, priceless.
Well, in simple terms, Spotify found a way to make these mix tapes (now called Playlists) for you, at scale. Of course, you can still hand-pick your own songs, create your own labels and digitally share or gift the playlist.
But wait, it gets better and the real magic happens when Spotify predicts what you would like listening to next and feeds you a continuous stream of music that not only includes your favorite hits but also new music that is different yet familiar, helping you discover and love those songs you would have otherwise not. Do you want to know how? Let’s look under the hood.
Any great digital experience is a harmonious balance between ‘Science’ and ‘Art’ and the Spotify experience is no different. Here’s my attempt at decoding this mix of Science and Art and applying the same principles to a re-imagined shopping experience.
A. The Science behind Spotify’s Personalization
Spotify uses something called an “Algatorial” approach to recommendations. Yes, you guessed it right — it’s a made-up word, which is created by combining two words — Algorithms and Editorial. Let me explain.
While Algorithms refer to the rules that computing systems use to generate a list of songs you will love, Editorial refers to the human touch, by which a team of in-house experts curates playlists based on genres, artists, and moods.
It was very clear to the Spotify team early on that while the editorialized playlists were a great way to get started, they were not scalable to meet the diverse tastes of its 100M+ users. Therefore, it needed a set of algorithms and through experimentation and continuous learning, it has evolved into using the following types:
1. Content-based Personalization
The goal here is to really decode the type of content (in this case songs) you listen to and convert it into values (aka Song Vectors) that can later be used to find more songs, with similar values, that would match your taste (again measured in values aka User Vectors). This is done by the following types of analysis:
a. Metadata analysis using Natural Language Processing (NLP)
This is primarily the analysis of the text associated with each song acquired from various sources including name, artist, description, genre, news, review, blogs or another type of content written on the internet.
b. Audio Analysis using Convolutional Neural Networks (CNN)
Metadata analysis works well when there is enough content available for a song, but what about when the song or the artist is relatively new? How would you find similar songs or people who are likely to love this song?
This is where Audio Analysis using CNN steps in. This is the same technology used for facial recognition and uses an Audio Spectogram, which consists of several attributes such as frequencies, beats, melody, lyrics, etc. that are decoded and converted into values (Latent Space Vectors) to then be matched against other songs and users.
Now let’s try to apply these principles to fashion. For example, if we decode a dress instead of a song, this is how it would look:
2. User-based Personalization
a. User behavior analysis using Collaborative Filtering:
When you as a user start the journey with Spotify, there is very little data to start serving you extremely accurate recommendations. But thankfully, we as humans are more homogenous than we may think. This means that our music tastes are shared across a large number of people, who can be clustered into groups. Further, if you can build a matrix of all these users within a cluster who share songs they love, you can start filling in the blanks using songs that some of the users love but others within the same group haven’t discovered yet. This is essentially collaborative filtering. You then enter into the learn 👉 improve 👉 iterate cycle.
B. Art of Personalization at Spotify
This is where you have to move on from the math and focus on the human element. I can break down this area of Spotify’s personalization into 3 key pieces:
a. Context (Situation)
In this case, Spotify scores very highly in opinion, since they have really tried to tailor user experience based on where and how consumers use their app. For example, you would see suggestions for different playlists based on time of day, day of the week, season or holiday, or even based on whether you are on move. By personalizing the experience given the context of the user experience, it creates a magical micro-moment for the customer that is hard to forget.
b. Design (UI / UX)
It’s a cliche that simplicity is the ultimate sophistication. But best designs usually are simple and I think Spotify gets quite close. Navigating through the app is quite simple yet full of surprises thanks to its ever-changing content.
c. Editorial (Curation)
Finally, if you are running out of ideas on what music to listen to, then there is an editorialized section that has curated playlists for each occasion, genre, mood, or situation.
Bringing it all together to re-imagine Spotify as a shopping app
To do this exercise, I created a framework of personalization with the same underlying principles of Science and Art I described earlier; Let’s call it True Personalization. With this framework, I’m proposing that a truly personalized experience must not only be personal but also relevant and contextual.
Now let’s apply this framework combined with the principles of Spotify experience to create a personalized shopping experience.
A. PERSONAL: let’s start with making it Personal by looking at how Spotify makes it happen.
This is how an equivalent PERSONAL shopping experience would look like.
As you can see from the two set of images above, just as Spotify allows you to quickly access the songs that matter to you based on your interactions, so should you be able to access the products that you recently viewed, searched for, or even added to cart to pick up your journey from where you left off. This experience clearly reflects your personal digital footprints on the platform and begins, what is really Step 1 in the Personalization journey.
Next, now that we have some clues about your preferences, we need to tailor your discovery process to make it more Relevant for you.
B. Relevant: this is where the Science kicks in and we need to devise a way to find more relevant products for you based on the Content and User-based algorithms described earlier.
Here’s how we could create a more relevant experience based on your activity and any data we can collect about you — both implicit: gender and interests obtained from Facebook in case you logged in using this service and explicit: items you browsed, added to cart / wish list or purchased. See some example below:
C. Context: next, let’s look at how using the data we have thus far, we can add contextual information and further enhance your experience.
The way Spotify adds context is so simple and intuitive, you wonder why others don’t do the same. For example, I get “start the week strong” recommendations on Mondays and “the weekend is here” on a weekend. Also, interestingly, while using Spotify on the Car-Play in my car, I started getting recommendations for my commute. Clever and so timely.
D. True Personalization: finally, the magic with Spotify happens when it brings all this information together to deliver a unique experience for you, which is a mix of both familiar and new. There are some key insights worth noting here.
- The Discover Weekly playlist is a compilation of new songs you have never listened to before but are based on the songs you have listened to and loved. Even though I had not heard of some of the artists or songs on that list, they seemed quite familiar and quickly moved to my favorite playlists. I was surprised to learn that I have a thing for Pop, especially some of the songs teens are listening to (embarrassing, but true 😳). If it wasn’t for the algorithm, I would have never found those songs in my library.
2. The Daily Mixes: this blew me away the first time I discovered this personalized playlist. After my first two weeks, all the songs I had heard were neatly organized into playlists based on the genre of music — Western Pop, Indian, Latin, Electronic, and Meditation Music. So whatever mood I was in, I had a playlist for it. Remember, mix tapes? 😃
3. The Surprise element: As the year came to a close, I was presented with insights into my listening habits with my favorite songs of the year, number of hours of music I listened to, number of times I listened to my favorite song, a comparison of my tastes with others in my age group, etc. It was a great use of data to educate and surprise me, and a friendly reminder that someone was working hard to make sense of all my data. While some may feel worried about privacy, I absolutely loved it.
Applying the same principles as listed above, you can create an equivalent fashion version. See below:
While e-retailers have long aspired to personalize the customer experience, very few have come close to one that quite matches up to the Spotify experience. This exercise has helped me better understand the fundamental principles and has inspired me to work harder to deliver delightful customer experiences through automated and scalable personalization. The abundance of data and computing, and the sophistication of algorithms are opening up new horizons for personalization and I can’t wait to see and help shape the future.