Netflix/Pandora and Neural Networks

[steppinthrax]

I want to know (if someone can assist) how do media services like Netflix, Pandora craft and profile music/movies based on people's taste.

I've tried to find info on this topic, however I'm looking for something highly technical and some code implementation of how this works (generically).

My thought is when you "thumbs up" a song on Pandora, you are "averaged" against other's who also thumb up the same song. Then you are thrown in a "song pool" of people who also like that song as well as others, therefore you hear other songs they also thumb up....

As you thumb up more and more songs people are removed from your song pool. While your song pool gets more specialized.

Not sure if this is how it works, just my wild ass guess.

 

[Biftheunderstudy]

While Netflix almost certainly uses something a bit more complicated. Try:
https://en.wikipedia.org/wiki/Collaborative_filtering

 

[Cerb]

On Pandora, you can easily tell by experimentation that thumbs-upping a lot expands the range. Down-voting narrows it. It's kind of counter-intuitive. Downvoting the same artists twice removes it, either permanently or for some weeks (you can use add variety to get them back). They use the various properties in their DB(s) from processing songs to link songs and artists, for a given station.

By their description, they use properties of the song(s) you add (if you use an artist, they use some collection of their songs) and thumb-up to figure out others. You will find common descriptive attributes amongst your songs, if you check the info, FI.

AFAICT, Pandora does not use the kind of collaborative suggestions you're talking about, though I wish they did. While I've gotten to where I can train my stations, now, using the technical properties of the songs often misses the mark, when it comes to what makes something good or bad, as a listener.

 

[William Gaatjes]

Since i am a "out of the box" kind of guy, i get generally annoyed by view recommendations. Those algorithms work by grace of that when you take a group of people, most will have similar interests. When you do not fit into that group, it does not work.

For years, youtube has recommendations for me that i do not like at all. But since my music choice is so diverse, the recommendation algorithms just fail to grasp what i like. With netflix it is the same. All those series and movies that i get recommended to, 9 out of 10 i do not care about. Unfortunately for me, the movies they do get right, i have in my own movie collection already...

Oh well, i will just wait until all the big bang theory episodes are available.

 

[sm625]

You dont want to use Netflix's code. It isnt about creating a profile that matches your tastes. It is about making money, and little more. Netflix will deliberately steer you away from high end content. In fact there have been many times where I have typed a movie into Netflix search, and had it return no results. But there was a movie that I knew was on Netflix. I knew because I read an article saying it was available to stream, and that is why I opened netflix in the first place. I literally had to google the movie's name followed by "netflix" in order to get a valid link to the movie. After I did this, I went back to my netflix home page, and then typed the name of the movie again, and this time it showed up in the results. This has happened several times. I think the reason it happens is because netflix must have to pay based on how many times people watch certain movies. This is why they attempt to steer you away from content that costs them an arm and a leg. A company that does this sort of thing cannot possibly give an honest assessment of your tastes.

 

[William Gaatjes]

You dont want to use Netflix's code. It isnt about creating a profile that matches your tastes. It is about making money, and little more. Netflix will deliberately steer you away from high end content. In fact there have been many times where I have typed a movie into Netflix search, and had it return no results. But there was a movie that I knew was on Netflix. I knew because I read an article saying it was available to stream, and that is why I opened netflix in the first place. I literally had to google the movie's name followed by "netflix" in order to get a valid link to the movie. After I did this, I went back to my netflix home page, and then typed the name of the movie again, and this time it showed up in the results. This has happened several times. I think the reason it happens is because netflix must have to pay based on how many times people watch certain movies. This is why they attempt to steer you away from content that costs them an arm and a leg. A company that does this sort of thing cannot possibly give an honest assessment of your tastes.

Their way of presenting is a bit strange. And it seems their searchmachine is a bit off. I do find it sad i can not just scroll through their complete catalogue.

 

[John Connor]

Pandora hasn't really worked for me if that's what they do. But if they are I'm sure it's based on musical qualities like rhythm, bass, certain influences, a vocal harmony, etc, etc, etc...

 

[Cerb]

Pandora hasn't really worked for me if that's what they do. But if they are I'm sure it's based on musical qualities like rhythm, bass, certain influences, a vocal harmony, etc, etc, etc...

Since this thread, I've been experimenting a bit with it (to see what I can make it play in odd cases, rather than just training it to play less crap I don't want), and it doesn't seem to handle influences at all, merely the pure processed metrics from the songs.

FI, it's not hard to use downvoting of other rappers to get Johnny Cash to come up a lot on a Missy Elliot station. If you remove the context, they have a lot in common, when it comes to timing and tonality. Likewise, it's basically impossible to seed a channel with influential artists, and get more than one or two of those they heavily influenced, with those coming up sounding very similar. You really have to use both the influencers and influencees as seeds, to get anywhere; because it goes by, "what they sound like," rather than, "how they sound."

 

[William Gaatjes]

I did noticed something funny and disturbing at the same time, today with youtube.

I have the link for this song :

Beastie Boys - Multilateral Nuclear Disarmament

https://www.youtube.com/watch?v=r3W7koH-isE

While playing the song, i closed down the browser for a flash update.
Afterwards, i start my browser again and instead of youtube loading Beastie Boys - Multilateral Nuclear Disarmament
, Beasty boys - Intergalactic was loaded.

I do hope this is just a fluke, but if this is going to be a trend then it seems that in the near future, youtube links may not link to the song you think you are hoping to get. Just the ones deemed popular by the mass voting.

 

[John Connor]

Since this thread, I've been experimenting a bit with it (to see what I can make it play in odd cases, rather than just training it to play less crap I don't want), and it doesn't seem to handle influences at all, merely the pure processed metrics from the songs.

FI, it's not hard to use downvoting of other rappers to get Johnny Cash to come up a lot on a Missy Elliot station. If you remove the context, they have a lot in common, when it comes to timing and tonality. Likewise, it's basically impossible to seed a channel with influential artists, and get more than one or two of those they heavily influenced, with those coming up sounding very similar. You really have to use both the influencers and influencees as seeds, to get anywhere; because it goes by, "what they sound like," rather than, "how they sound."

You can add more than one station to a station to get better variety.

 

[Cerb]

You can add more than one station to a station to get better variety.

I do. That's one of the things thing I haven't found on Spotify's radio feature, that keeps me using Pandora, despite the annoying skip limits (another is that Pandora seems to have normalized ads, whereas a Spotify ad might jolt me out of my seat - the latest Kenny Chesney(sp) one as to be 10+dB louder than the metal I'm trying to listen to!). I can have a station with The Birthday Massacre, Billy Preston, Alro Guthrie, and Rob Zombie all in the list of seeds, and get some nice variety, over time, within a single station.

 

[Cerb]

I do hope this is just a fluke, but if this is going to be a trend then it seems that in the near future, youtube links may not link to the song you think you are hoping to get. Just the ones deemed popular by the mass voting.

Look to the right of the video. There's an autoplay switch. Turning it off should prevent that from happening in the future. By default, it will load the first thing in the right-hand-side recommended list, I believe.

While I don't know for sure, given the results on more fringe artists, and technical subjects, I suspect that Google uses clicking to another video and playing it for the recommendations, along with the voting.

 

[William Gaatjes]

Aha, that is likely possible. But when i clicked the link again i got the correct video. So, not the first of the recommended list.

I did notice with the netflix app, that claims i have watched movies that i have not. Not even clicked them. And the recommended list depends on what i have seen. I wonder if they have local streaming servers per country. If not, could be fun if someone found a way besides vpn to get more movies.