Spotify's music recommender algorithm: How it works
ฝัง
- เผยแพร่เมื่อ 22 ต.ค. 2022
- From the SDS 619: Tools for Deploying Data Models into Production, where @JonKrohnLearns speaks with Erik Bernhardsson, the man who invented Spotify’s original music recommendation system. They address the different ways to interview a data science candidate, how to deploy a data model into the cloud, and the approach he took that made Spotify go from a digital music startup to an AI-driven streaming giant.
Watch, listen to, or read the full episode at www.superdatascience.com/619 - วิทยาศาสตร์และเทคโนโลยี
Even though the Spotify algorithm gets a lot of hype for being super advanced, a ton of folks on those big forum sites where music is discussed seem to think it's a total letdown. According to them, Spotify's algorithm just sucks. But guess what? They've found a real gem that beats it hands down: TH-cam Music. Apparently, TH-cam Music's algorithm works like pure magic and blows Spotify out of the water. It's way, way, way better, no contest. So, like, nowadays Spotify is all about getting people to talk about their algorithm and how amazing it is. They even pay folks to do it, which is basically a big advertisement for the company. But let's get real, when it comes to actually listening to music, the fancier their algorithm gets, the less likely it is to play songs you actually enjoy. It ends up being more repetitive and you find yourself skipping through more and more songs most of the time. It's like, what's the deal with that.
Cut down to 2 months Google fired the while YT Music team. 🤷♂
Then would it be a matter of being "too accurate" then? If a less accurate algorithm helps keep the listener entertained by adding in songs that the user isn't expecting, that same issue can probably be "fixed" by adding a randomness factor to add variety to the listener.
Thanks a lot for the insight from your guest. I was able to use quite a bit of the background information for my oral exam about personalization algorithms and the power that playlists hold! You also have a great way to listen to your guests and the way you ask stimulant questions is very natural :)
I found your channel previously and this channel is awesome, I hope you will get good achievements about this channel because it deserves
It's sort of like word embedding but where all the context of the song is taken from what users listened to
thank you i was so curious about it few videos really talks about it. there is more videos about “blowing up” as an artist on spotify
Nice! Glad we were able to provide you with what you were looking for!
Small critique: Dude, just listen to the guest! You are still writing and doing something with the laptop, and it looks like you are not paying any attention. You can make notes later on since you have it recorded.
this
Really insightful stuff!
For sure! Erik's brilliant.
My recommends is broken, it plays the same songs all the time LOL, even Shuffle
Try discover weekly playlist again and again.. recommondation will change accordingly..
Interesting, I was considering doing a master thesis in recommender systems. One of my motivations being how bad spotify is. TH-cam on the other has suggested tons of new music to me I never would have heard otherwise. But maybe Im one of the notorious grey/black sheep that is hard to make suggestions for
Spotify recommendations are so awful it truly boggles the mind. And that's after feeding them literal years of playlists and feedback via 'enhanced' playlist recommendations. I also know this experience happens to more users than it doesn't and even this sparse comments section confirms this. Classic example of optimizing for the wrong metric (whatever that may be). At least now I have a face to direct my disdain at
it works like crap lolol
It certainly can! Did you hear Erik's joke about this?