Christine Hung is Head of Data Solutions at Spotify. She spoke at Data Driven NYC on October 16th, 2017, about how the music streaming service is using data to delight their users with personalized playlists.
Below is an unedited transcript of Christine’s talk:
My name is Christine Hung and today I’m going to talk to you about music and Spotify. Before I begin I have to ask the obvious question, do you guys listen to music? Yes. Raise your hand if you listen to music. Okay some people didn’t raise their hands. Raise your hands if you do not listen to music. Great. We have a few. Often times, when I ask this question people are just a little reluctant to show their hands.
Hopefully I can convince you guys to at least try out Spotify at the end of this presentation. Speaking of listening, people are actually streaming a lot more. Like basically more than ever. Thanks to the wonderful technologies that’s been made available in your homes, on your laps and also in your pockets. At Spotify alone, we now have 150 million active monthly users streaming from more than 60 countries.
If you think about when Spotify started, we’ve been around for 11 years. So that’s a cumulative of 11 years of streaming data for us to analyze from. I believe that adds up to more than 30 terabytes of data. You might be wondering, why are people streaming so much? Aren’t we competing with the publishers, the social media companies, and we’re competing for people’s attention.
The interesting thing about music is that we’re actually not competing as much as you think and that’s because music is complimentary. If you think about your own listening behavior. Think about when you listen to music. You probably have music in the background when you were Tweeting, when you were on Facebook, you probably had music on when you’re working or when you are working out. Right, there are many different sort of occasions where you can have music but do something else at the same time.
According to research, we also know that a lot of people even listen to music when they’re sleeping. That’s the reason why music is actually, it’s basically like a mirror, and it reveals who you are whether you like it or not. That’s the reason why at Spotify, because we have so much data that we can use, we were able to continue to find good ways to improve your user experience, come up with personalized experiences like your discover weekly or your daily mix.
Today I’m going to use myself as an example to show you how this works. As you can imagine at Spotify, we talk about music a lot and it’s inevitable that people ask you what kind of music do you listen to or who are your favorite artists? If you ask me that question, this will be my official answer. I’m going to tell you that I’m really into Latin music. I really like Manuchao, who’s a famous Spanish punk rock artist. I’m also going to tell you that I really like Marisa Monte. She’s famous for Bossa nova, Brazilian artist.
I will also tell you that I like Jack Johnson a lot. Jack Johnson’s my go to when I need something just chill in the background. Of course Madonna. The fabulous, talented, [famless 00:03:40] icon. So it’s all good right because this sort of fits the image that I want for myself because at Spotify I want my colleagues that I’m a serious music fan and I know what I want, but if you look at what I actually listen to, it’s actually a very different story.
Let’s just look at what I stream the most for the last six months. Let’s look at the top three artists that I stream the most. Number three was Manuchao, I just told you about. He’s one of my favorite, which makes a lot of sense. This is the number two artist that I streamed the most. Does anyone know who this is? It’s not Britney Spears. It’s not Madonna. This is Paulina Rubio.
She’s a very famous Mexican pop singer, and according to one of my friends, she’s the Britney Spears of Latin music. This will be someone that I typically would not tell you about because she’s as main stream as it can be. Like I say, I do not want the image of me being affected by mainstream sort of top 100 list of top artists. So now you know how this goes. Can you guess who was my most streamed artist? Just guess.
Is it Justin Bieber? Or is it Taylor Swift? The reality is I don’t even know who it is. This is what I stream the most. Let’s just pause for a second and think about what’s happening. I stream this track 792 times in the last six months. So what does this mean? Let’s think about what does this mean. You can obviously you know in my example you can tell that perception is not reality. There are always these little secrets that we want to keep to ourselves, but I think this actually tells you a little bit more.
Obviously you can see that I’m pregnant and I’m having a baby, but do you think I stream that song for this baby in my belly so that she can remember the lyrics when she was born? Probably not right? You can probably confidently guess that I am already a mother and I have a toddler at home. At least one toddler at home whose obsessed with Wheels on the Bus. Right? Makes sense.
From this example you can see that because music reveals so much of ourselves, at Spotify, we are beginning to explore what we can learn about the audience just based on what they’re listening to. So I’ll give you a few examples. Can we measure how open people are when it comes to discovering music. Right? From there can we then infer okay because you are open to discovering music, you might also be open to trying out new products or trying out new experiences.
Can we understand better in terms of how diversified your listening taste is? Are you a loyalist who will stick to one genre or just a few artists, or are you more of an eclectic listener where you can just jump around between different things. Also, can we learn about how laying in or leaned back you are? Are you a curator who always knows what to play, or are you more like me?
I often times just want something in the background while I work. As we start to capture different types of user behavior, we are able to start mapping out critical moments of your life. Going back to my earlier example, right? If you look at my listening history, you’re going to see that I listen to Wheels on the Bus on repeat, almost every night between six and seven. Not tonight, but in general that’s what I do. You can probably guess that’s when I’m at home playing with my daughter and she’s obsessed with the song.
You will also see that I use my running playlist every weekend, Saturday, Sunday between exactly 8 and 8:30. What does that tell you about me? It’s probably pretty obvious that I also run during the weekend and that means that I’m probably into sports and sort of athletic products. I mean not right now, but that’s something that you can start to learn about me. You might wonder, so how do we do all of this?
It’s very simple. Just like I’m sure many of you, you’re building machine learning algorithms, you’re doing a lot of different prediction work, so at Spotify we spend a lot of time collecting ground tooth data. We primarily use the on-platform signals that we have. We then test it out. We tweak it and then we turn them into real products. Here’s an example of the type of testing that we do. If I have gamers in the audience, you guys probably know that about a month ago, we launched the integration with Xbox.
So with Xbox, with the launch, the biggest question in our marketing team’s mind was that okay so how do we find these gamers. So this time, instead of going with sort of the traditional route if you think about who are the gamers. If you think about demographics, they tend to skew more towards male. They tend to be younger, but instead of doing that, we decided to just go straight to our modeling and predict directly who the gamers will be. This is the result that we saw.
We saw a significant increase in cross-device usage, which is what you always want. We saw a much higher engagement rate. We were able to dramatically reduce the email up dial rates as well because the messaging was really relevant. Most importantly, we were able to validate these new segments that we’ve been creating and from there you know, continue to iterate.
I thought before I run out of time, I can just talk about one more thing. If you can all just follow me and close your eyes just for a little bit, won’t be too long. Close your eyes and try to think of a time when you heard a song from TV or on the radio station that triggered some interesting, distant memories for you. It could be the song from the first dance of your wedding. It could be a song that you listen to when you went on a road trip.
Or it could be a song that someone special sent you in a mix tape. Now think about how powerful that feeling is. Okay, now you can open your eyes. Thank you for cooperating. The feeling that you just had is called nostalgia. If you ask me what I’m most nostalgic of, and this time I will be honest, what triggers very strong memories for me, it would be the Backstreet Boys. It would be Spice Girls, and then of course it’s going to be Britney Spears.
Now you probably know how old I am. Especially for those of you who are in your mid 30’s, you probably feel very strongly about these artists too. To me, music is a way to remember, right? Those artists bring me back to my high school, my college days, where I was so focused on dancing and exploring the world. At Spotify, we spent the last 12 months studying your nostalgia, and for those of you who use Spotify, you probably know that about a few weeks ago we just launched this new product called Your Time Capsule.
This is basically a personalized playlist tailored specifically for your nostalgic obsession. Let’s talk a little bit about how we came up with this. We believe that nostalgia is a very strong emotion. So what we thought a year ago when we started this journey, we thought okay let’s just look at what nostalgia actually means. On this chart, what you see is on the Y axis, you see basically all of our users current age.
At the X axis what you see is the age of the users when the music that they listened to actually came out. The color represents intensity. You can obviously see that there’s a pretty high concentration here. The way to interpret this chart is that regardless of how old you are, at least on Spotify, people keep going back to the music that they came out during their high school or during their teenage years. Which is pretty powerful. You would think okay, how do we come up with this playlist?
Do you just look at how old people are and we just give you your high school playlist? Of course not, of course it’s a lot more sophisticated than that. We actually ran a lot of tests internally just with employees to see how we get it right. Basically here are the three major tests that we did. We first look at, so on Spotify it’s very easy for us to identify the songs that make you feel nostalgic. That’s the basic list that we provided.
We basically grouped that based on your age, and your country. We also tested out sort of, basically the billboard like the top songs, right? That’s another category. Of course, we have to use our machine learning model. We also did another experiment that’s just basically a combination of the first two, but we really customize to your tastes based on what you had been listening to. Guess which one is the most successful?
Obviously it’s the third one. That’s how we came up with Your Time Capsule. I believe this is all the time I have today, but I’m happy to chat more and feel free to ask me questions. The only thing I would say is if you have not tried Your Time Capsule, please try it out and let us know how you like it. Thank you.
Long Term Memory for AI with Pinecone’s Edo Liberty
A Conversation with Chris Wiggins, Author of “How Data Happened”
An AI Assistant to Work Faster with Notion’s Daniel Sternberg
Entering the Data-Centric Era of Foundational Models with Snorkel AI’s Alexander Ratner
LLM Powered Search with Vectara Founder & CEO, Amr Awadallah
04.18.23 / Amr Awadallah, Vectara
Build and Deploy AI with Pytorch, a conversation with Lighting AI Founder & CEO, William Falcon
04.18.23 / William Falcon, Lighting AI
Generative AI for Speech Recognition with AssemblyAI Founder & CEO, Dylan Fox
04.18.23 / Dylan Fox, AssemblyAI
A Conversation with Sarah Catanzaro from Amplify Partners
03.15.23 / Sarah Catanzaro, Amplify Partners
The 2023 MAD (Machine Learning, Artificial Intelligence & Data) Landscape
03.16.23 / Matt Turck & Kevin Zhang, FirstMark