In what universe could this information possibly lead to an improvement in recommendations? You can't derive subjective taste preferences from an emotional state - some people listen to metal when...
In what universe could this information possibly lead to an improvement in recommendations? You can't derive subjective taste preferences from an emotional state - some people listen to metal when they are angry, others listen to classical or pop. This seems... insane, divorced from the reality of music appreciation. Gender has no role - I know kids who listen to progressive rock and old people who listen to black metal and women who listen to the worst kind of hip-hop. Is this just a pretty patent to lure dumb investors?
If you want good recommendations, there's only one way to get them - ask people, not algorithms. They'd have better luck tracking viral spikes in various tracks/artists and amplifying them to reach more people, or by improving the system so that users can more easily curate and share recommendations with each other.
A big problem I think is that Spotify has no context of like. Song similarity? It just has "User likes this" or "User doesn't like this." There's no concept of "User isn't in the mood for this...
A big problem I think is that Spotify has no context of like. Song similarity? It just has "User likes this" or "User doesn't like this." There's no concept of "User isn't in the mood for this right now." And the feature that would seem to be "Find more like this" is far too heavily weighted by that mistaken "Will user like this?" prediction.
The users are telling Spotify what music they want already. IMO what Spotify needs to do is start mining their music for more info, not their users. Even 20 years ago, the PSP and Sony's SensMe channels were extracting mood and energy info from music and splitting it up into Spotify-style mood playlists entirely automatically. As a recommendation engine it was better than anything Spotify is doing now. And it was completely private and local-device-only, capable of running on a sub-GHz MIPS processor.
Yeah totally! I’ve been a little tuned out of the music information retrieval world lately, but this is a pretty common subject of research. Here’s some papers from past ISMIR (International...
Yeah totally! I’ve been a little tuned out of the music information retrieval world lately, but this is a pretty common subject of research. Here’s some papers from past ISMIR (International Society for Music Information Retrieval) conferences:
Not that I've found but I've always wanted to experiment with it. I feel like I at least know enough from making music visualizers to find out how to shoot myself in the foot very effectively...
Not that I've found but I've always wanted to experiment with it. I feel like I at least know enough from making music visualizers to find out how to shoot myself in the foot very effectively here. Sadly it's hard to find the time for a project like that.
Agree on the time front. I did find multiple nice papers and couple of semi-useful implementations. But creating the feedback loop and generating a self-updating local database is a challenge for...
Agree on the time front.
I did find multiple nice papers and couple of semi-useful implementations. But creating the feedback loop and generating a self-updating local database is a challenge for me as it is. I would like it truly global, ala MusicBrainz, but that is a lofty undertaking.
Spotify isn't interested in recommending music you'll think is good. It's interested in recommending music that will get it the most money. In the short term, the loss of revenue from those with...
If you want good recommendations
Spotify isn't interested in recommending music you'll think is good. It's interested in recommending music that will get it the most money. In the short term, the loss of revenue from those with non-mainstream tastes will be outweighed by the vast majority of users who mindlessly let it play whatever it pushes at them, which mostly will be the songs that fall under its new revenue-sharing model where artists get less money per play and Spotify gets more.
Oh sure, I can't say for sure that's what's going on, but my prediction is based on a pattern of businesses doing similar things in the past, and I've been given no reason to believe they'll...
Oh sure, I can't say for sure that's what's going on, but my prediction is based on a pattern of businesses doing similar things in the past, and I've been given no reason to believe they'll behave differently.
On some level it's not entirely their fault. The world of music licensing is one of the finest nightmares ever created by capitalism, and if you have to deal with labels and the business side of...
On some level it's not entirely their fault. The world of music licensing is one of the finest nightmares ever created by capitalism, and if you have to deal with labels and the business side of the music world, you're guaranteed to land far from an optimal solution. Spotify can't change the system, they just have to live with it.
While you're certainly not wrong about what a nightmare music licensing is, they don't have to just live with it. They could do what similar companies like Netflix, Amazon, etc. do and make their...
While you're certainly not wrong about what a nightmare music licensing is, they don't have to just live with it. They could do what similar companies like Netflix, Amazon, etc. do and make their own content in addition to licensing others, and divide the profits more fairly. They'd likely make more and so would artists because there'd be no middle man. It's not black and white. There are other alternatives.
Maybe it is not used strictly for recommendation, but rather to identify the speakers preferences and create a playlist that mixes their already determined preferences on the fly?
Maybe it is not used strictly for recommendation, but rather to identify the speakers preferences and create a playlist that mixes their already determined preferences on the fly?
Why go the invasive route for this? You can ask your customers. But even with much stronger indications of music preferences (a very long list of likes), Spotify still kinda sucks at recommending...
Why go the invasive route for this? You can ask your customers. But even with much stronger indications of music preferences (a very long list of likes), Spotify still kinda sucks at recommending music. All my song radios are basically remixes of my liked playlist, no new songs and hardly any of them are actually similar to the song I requested anymore. This new info would be more difficult to use and less accurate to boot.
I don't understand how this technology could benefit Spotify's users. But I definitely understand how this could benefit Spotify. I wonder what the implications would be for a system that plays...
I don't understand how this technology could benefit Spotify's users. But I definitely understand how this could benefit Spotify.
I wonder what the implications would be for a system that plays 'random' music for you while monitoring your camera and microphone to gauge your reaction then applies an AI system to parse out useful data.
I feel like I would actually give that a try and be 100x more comfortable (assuming you could turn the feature off) with versus what this article is describing.
In what universe could this information possibly lead to an improvement in recommendations? You can't derive subjective taste preferences from an emotional state - some people listen to metal when they are angry, others listen to classical or pop. This seems... insane, divorced from the reality of music appreciation. Gender has no role - I know kids who listen to progressive rock and old people who listen to black metal and women who listen to the worst kind of hip-hop. Is this just a pretty patent to lure dumb investors?
If you want good recommendations, there's only one way to get them - ask people, not algorithms. They'd have better luck tracking viral spikes in various tracks/artists and amplifying them to reach more people, or by improving the system so that users can more easily curate and share recommendations with each other.
A big problem I think is that Spotify has no context of like. Song similarity? It just has "User likes this" or "User doesn't like this." There's no concept of "User isn't in the mood for this right now." And the feature that would seem to be "Find more like this" is far too heavily weighted by that mistaken "Will user like this?" prediction.
The users are telling Spotify what music they want already. IMO what Spotify needs to do is start mining their music for more info, not their users. Even 20 years ago, the PSP and Sony's SensMe channels were extracting mood and energy info from music and splitting it up into Spotify-style mood playlists entirely automatically. As a recommendation engine it was better than anything Spotify is doing now. And it was completely private and local-device-only, capable of running on a sub-GHz MIPS processor.
Do we have any open source implementation for mining mood data from music?
Yeah totally! I’ve been a little tuned out of the music information retrieval world lately, but this is a pretty common subject of research. Here’s some papers from past ISMIR (International Society for Music Information Retrieval) conferences:
http://ismir2011.ismir.net/papers/PS6-18.pdf
http://ismir2009.ismir.net/proceedings/OS5-4.pdf
http://ismir2006.ismir.net/PAPERS/ISMIR06105_Paper.pdf
Not that I've found but I've always wanted to experiment with it. I feel like I at least know enough from making music visualizers to find out how to shoot myself in the foot very effectively here. Sadly it's hard to find the time for a project like that.
Agree on the time front.
I did find multiple nice papers and couple of semi-useful implementations. But creating the feedback loop and generating a self-updating local database is a challenge for me as it is. I would like it truly global, ala MusicBrainz, but that is a lofty undertaking.
Spotify isn't interested in recommending music you'll think is good. It's interested in recommending music that will get it the most money. In the short term, the loss of revenue from those with non-mainstream tastes will be outweighed by the vast majority of users who mindlessly let it play whatever it pushes at them, which mostly will be the songs that fall under its new revenue-sharing model where artists get less money per play and Spotify gets more.
Considering that we have no insider knowledge how the algorithm works, you seem overly confident about this.
Oh sure, I can't say for sure that's what's going on, but my prediction is based on a pattern of businesses doing similar things in the past, and I've been given no reason to believe they'll behave differently.
On some level it's not entirely their fault. The world of music licensing is one of the finest nightmares ever created by capitalism, and if you have to deal with labels and the business side of the music world, you're guaranteed to land far from an optimal solution. Spotify can't change the system, they just have to live with it.
While you're certainly not wrong about what a nightmare music licensing is, they don't have to just live with it. They could do what similar companies like Netflix, Amazon, etc. do and make their own content in addition to licensing others, and divide the profits more fairly. They'd likely make more and so would artists because there'd be no middle man. It's not black and white. There are other alternatives.
Maybe it is not used strictly for recommendation, but rather to identify the speakers preferences and create a playlist that mixes their already determined preferences on the fly?
Why go the invasive route for this? You can ask your customers. But even with much stronger indications of music preferences (a very long list of likes), Spotify still kinda sucks at recommending music. All my song radios are basically remixes of my liked playlist, no new songs and hardly any of them are actually similar to the song I requested anymore. This new info would be more difficult to use and less accurate to boot.
I don't understand how this technology could benefit Spotify's users. But I definitely understand how this could benefit Spotify.
I wonder what the implications would be for a system that plays 'random' music for you while monitoring your camera and microphone to gauge your reaction then applies an AI system to parse out useful data.
I feel like I would actually give that a try and be 100x more comfortable (assuming you could turn the feature off) with versus what this article is describing.