How the Algorithm actually works (mathematically explained)

[@gaiomuzio]

Luckily the algorithm recommends me your videos

[@some_one]

I thought this was gonna be about an algorithm for matrix operations

[@t0maz.m]

Me too😢

[@deepslates]

I’m really interested in this, I’m working on a big project about social media algorithms and this was something I hadn’t seen before. Could you make a more detailed explanation or share somewhere where they explain it?

[@TheUnqualifiedTutor]

"Deep Neural Networks for KZbin Recommendations" (Covington et al., 2016)

[@deepslates]

@ Thank you so much.

[@Max_bibikov]

Great video lad!

[@STA-3]

I remember breaking the KZbin algorithm once, but i just guess it was a fluke 😭

[@aa6eheia156]

Wow 1 million+ views

[@STA-3]

@@aa6eheia156 Yeah... it's kinda crazy

[@808lilglo6]

I got that video in my feed too cuz i also make musik

[@STA-3]

@@808lilglo6 No way 😭

[@pez1870]

​@@STA-3i watched that video like 8 times in loop 😭

[@justice32legends]

So, what's the algorithm for streaming music, the same one? Or do they focus more on replays than playtime?

[@TriglycerideBeware]

Not the exact same, but very similar. It tracks a bunch of user behavior, like: playlists, listen time, skips, re-listen frequency, shares, favorite artists, favorite genres, etc. and compares it to other user behavior. A bunch of the metadata they use for comparing similarity of songs is harmonic information (that's how they know things like favorite "genre" like I mentioned earlier). There are some basic ones like tempo, key, chord progression, lyrics, etc., but there are also some more abstract metrics like danceability, energy, and acoustic-ness. It's much easier to auto-generate metrics like that for audio data than video data, though not impossible for video data. So it's overall still a combination of tracking user behavior to profile users' interests, analyzing media data and metadata to extract information in a format that is understandable to an algorithm, and then using an algorithm to map encoded content to encoded user profiles. The only differences are the specifics, e.g. the exact information encoded for the algorithm to use, and the exact algorithm. I don't know exactly which metrics are prioritized, but there is a bunch of information out there about the Spotify algorithm (like in dev blogs) if you want to know more.

[@justice32legends]

@@TriglycerideBeware You have my thanks for the elaborate explanation.

[@suzalisilunisenpro]

Wait so you use know how it works so you maximized you views 🤔

[@miner1546]

So "im not interested" Button doesnt does a thing then?

[@TheUnqualifiedTutor]

It probably adjusts your preference values

[@mladengavrilovic8014]

Can you make a more in depth explanation on the algorithm and explain how it tweaks the arrays?