Welcome back. It is my pleasure today to bring you an interview with Paul Lamere, who is the Director of Developer Platform at Spotify, one of the well-known online music companies. And who has been working in the area of recommending music for pretty much as long as people have been doing work in recommender systems. And I'm delighted to have you with us today, Paul. >> Well, thanks for having me. >> So we've spent a lot of time in this series of courses talking about recommending, largely from the perspective of examples like products generally, sometimes news, movies. What's different about recommending music? >> Well, there's quite a lot of things that are different about music. Probably one of the biggest things is that people listen to music over and over again. So lots of people will have their favorite playlists, and every day they're just listening to that playlist. And so these people are often not really looking for a recommender necessarily to help them find new music, but to help them to improve their listening experience. So one of things that we do in Spotify is try to nudge these non-adventurous listeners to expand their horizons a little bit, so that they can enjoy the millions of artist that we have in the catalog. >> So interesting, so how does this reconsumption change the way you think about recommendation? Obviously, I mean most of the algorithms we have taught start by saying, we're going to find something we think you haven't experienced. But I get the sense its a little more sophisticated than just saying great, we can now recommend the things you rated highly. >> That's right, yeah. So at Spotify, we think a lot about the user. And we actually have bucketed users into maybe three or four different types. At one end of the spectrum, you have the hardcore music listener who spends a lot of their recreational time exploring for music. And at the other end of the spectrum is you have the casual. Or maybe even people who if music want away they would not even notice or care too much, who spend most of there time traditionally listening to music on the radio. So they essentially had a zero button interface. They left discovery totally up to the DJ and they weren't paying to much attention to what they're listening. So the bottom line is, there's no one perfect recommender for all of the people who use Spotify. So especially for these casual listeners, giving them a good listening experience by helping them re-experience the music that they already have Is important. So what we have ended up doing is building a set of different recommenders. So for the hard-core music fan who's looking for music discovery, we have something called Discover Weekly. And every week you get a list of 50 songs by artists you've probably never heard of. Or at the other end the spectrum, we have something called Daily Mix. Which essentially takes the music that we already know that you know and love, and maybe every fourth or fifth song will mix in something new, but something that's not too far away from something that you already know. So the idea here is we're trying build recommenders that serve all these different types of users. >> So within this, are you trying to model somebody's frequency of re-listening, either as an individual or at the song level? And I know I can remember, as somebody who doesn't listen to a huge amount of music, there were certain songs that you would hear and you'd realize even a radio station might play it twice an hour because people wanted to hear that a lot. But they didn't want to hear everything a lot. And some people probably got really annoyed. >> Yeah, [CROSSTALK] some studies looking at terrestrial radio. We have a metric called the Time to Katy Perry- >> [LAUGH] >> Which is on certain commercial radio stations, the average Time to Katy Perry is about 30 minutes that you'll have to wait for a Katy Perry song to come up. So yeah, certainly, and actually terrestrial radio provides a pretty interesting model that we can look at. because they've been playing music for people getting close to 100 years now. And they have a lot of experience with mixing in familiar music and new music. And generally speaking, they've found that repeating music is a great way to keep people listening. So yeah, we use a lot of that type of experience to model our recommenders that we're building for this casual music listener. >> And do you also do something particular to think about, is there a song to song transition, whether one song follows another? Or is it just more, are they part of the same list? >> Yes, it depends on the context. So we have a lot of technology that lets us do things to improve song transitions. So we have an acoustic analyzer that can look at an MP3 file and give us the tempo and the energy, and how acoustic it is and how danceable it is. And for a certain context like dance music or for running, song transitions are very important. So we have some parts of our product that will use that data to give you seamless transitions, or to give you a playlist that has sort of a traditional DJ arc to it. Where it might start slow, and then from song to song gradually ramp up in energy until you get kind of the everyone out on the floor dancing. Then you may have a few cool off songs, and then start the whole thing over again. >> Wow, so things get pretty sophisticated. So when you think about this, and we've been over the course of a series of courses here, taking people from very basic non-personalized recommendation through content, user-user/item-item collaborative, and then into some of the matrix techniques. Are there particular innovations in recommender systems generally that were particularly helpful for making music recommendations effective? >> Yeah, so the things that we are spending a lot of time now using some of the Word2vec models. Are you familiar with those? >> Sure. >> Google's been working on it. So we have one of our great data sources, our user created playlist. So people have on Spotify have created about 2 billion playlists. And these are really great sources of information about how people actually listen to music. Because these playlists have titles and they have sets of songs that people think go well together. So we can toss all these into Word2vec model and essentially position all of these songs in a giant vector space. And then we could just do nice, nearest neighbors kind of things for song recommendation. So we end up leveraging this really large user-built data to give us a really good song to song similarity. And that forms the basis for lots of our recommendations. But of course, the trick is then taking that similarity and then making a good listening experience out of it. Fixing in the familiar, the new, the serendipitous surprise, and doing some of the things you were talking about with trying to give good song transitions and all that. >> Yeah, it's almost scary. I remember reading an article, I think it was in the New Yorker in the past year, about people on the music creation side who are using some of these same sets of technologies to try to engineer the songs to fit in, and be sort of instant hits. And the whole idea that, well gee, if we know what a recommender is going to think goes with a set of listening tastes, and we know what the tastes that are most popular are, then maybe we can engineer the beat and the lyrics to fit in with what somebody wants to listen to. And I don't know, for me at least there's some mix here between art and scary. >> Yeah, I'm certainly a skeptic about that sort of thing. There's been, especially in the music space, there have been a number of companies who have tried to build hit predictors. And [LAUGH] I think that's a little bit of snake oil. >> Yeah, I think it probably is a lot harder to predict a hit than to predict something or engineer something that it's sort of going to at least hit the main stream at a mid-level. >> That's right. >> But this is true in the movie space too. Everyone's figured out that okay, Halloween 12 or whatever the next thing in a series is, will have a steady audience. It may not be a blockbuster, but you know that- >> Right. >> You can do worse than Star Trek 25. >> Right, so the funny thing about these hit predictors, they oftentimes will look like they're doing a pretty good job and they'll have a good track record. But especially ones that are driven off of acoustic data, it turns out that they often are able to, essentially classifying the producer of the music. And so certain producers will have certain mixing styles that are easily recognizable. And so music by a top producer oftentimes becomes a hit. So if you can recognize a producer, you've done that. But that probably is not a hit predictor, it's really just a producer classifier. >> No, and in fact it, the thing I was reading was more on the producer end of saying yeah, some of these great producers just know how to match to what people are looking for. But that may not require all the algorithms, that may be a good ear too. >> Yep. >> So is there anything else about recommending music that you think would be interesting to share? >> Yeah, so one of the biggest things that we've been thinking about, is contextual listening of music. So I think, unlike most other domains, music listening is highly contextual. So people will build playlists for working out. And then, build a completely different style playlist for studying, or another one for doing chores around the house, another one for a date night, another one when their friends are coming over. So identifying context, identifying music that works well in context, and then personalizing this music for you based on your taste, is something that we've been thinking long and hard about. So if we know that you tend to like 70s classic rock and you're going to the gym, we can give you some AC/DC and some Led Zeppelin. But if you're a hip-hop fan going to the gym, we maybe be giving you some Eminem. And if you happen to be a female 17-year-old going to the gym, maybe we'll give you some Justin Bieber, or Ali J, or something like that. >> Well, you've just scared me off from going to the gym for the rest of the week. >> [LAUGH] >> But I think the whole concept of having the system recognize that context, and I'm still old fashioned in that I have a whole bunch of music sitting on the computer. Which now sounds old fashioned with playlists for here's a jazz party or here's whatever. And my major selector is random. >> [LAUGH] >> But, it's clear that I'd get a much better exposure if I uploaded these things and said, okay, now you can add things [LAUGH] to my list. And if I got that next step of it recognizing, wow, I got this for context. That sounds pretty cool. >> Yeah, so our phones are these, becoming these perfect context identification machines. They know when you're in the car, they know when you're running. They know where you are, they know who your friends are. They know your calendar. So it's actually kind of scary when you think about how much your phone knows about you. And that's one of the things we think about. We don't want to cross the creepy line in recommenders to say hey, we think you're going on your first date, here's some music. That may be something that you don't want your music service to know about. So, but understanding this context and then taking advantage of it is a pretty important part of what we're doing. >> Pretty cool. And if somebody wants to go be innovative, now that we're getting more devices that are reading your heart rate at the same time. It notices you're in the car, it notices your heart rate slowing down, it's time to wake you up. [LAUGH] And we'll pick the right music for the circumstance. And that'll be cool stuff. Well, thank you so much. >> Yeah, it's been my pleasure. >> Wonderful, thank you. And that's going to end our interview here.