Finding Similar Music Using Matrix Factorization

I think that for me the approach in TFA would work, if it was applied to songs rather than artists. There are many artists that have a couple of songs I like and a lot of ones I don't care for. Artist-based recommendations go terribly wrong for me for this reason, I think.

Add vocal to this list:

There is a lot of great music I don't listen to because it comes with the performers problems, complaints, and political beliefs attached.

Some people don't care but for those of us who do it would be great if there was a filter for spoken words.

While it's common to use things like genre, year, musical characteristics, etc, social dimensions may be at least as valuable. Exanples are artists performeing at the same show, collaborating on a project, maintaining personal friendships, signing to the same record label, moving to another group, etc. Most of the music I've found and enjoyed were through these types of connections.

I think Spotify's "Discover Weekly" playlists get pretty close to this. I really, really look forward to each week's new playlist. I listen to a wide range of genres, and completely agree with you that most discovery tools try to pigeon hole you into one or two. Spotify seems to pick up on other associations. I wonder how much neural net stuff they're doing with their big data.

What you're describing is is similar how Allmusic categorises artists (categorised? I haven't used it a lot in the last few years, so I don't know how much it's changed). I used to find most of my new music via it. It was in no way algorithmic - it didn't make suggestions, it was just a case of manually following links (given that part of the pleasure comes from searching and discovering independently, no bad thing). It had genres, mood, influences, influencees, etc at the artist level at least. It would seem that Rovi's API could be better leveraged to provide good suggestions than it currently is, if all that metadata is still being attached (though there are things like Spotify's Discover Weekly that I've found to be surprisingly good which I assume use it).

[It would have been even better if that had been expanded to album/song level. As the sibling comment suggests, categorising by subject matter would be very useful for a not-insignificant number of people as well.]

I am quite interested in this topic. There is some research on that field. Some links I collected in the past:

http://thesis.flyingpudding.com/

http://dynampd.ubitux.fr/

http://erikbern.com/2015/09/22/presentations-about-spotify-m...

http://erikbern.com/2015/09/24/nearest-neighbor-methods-vect...

http://www.wired.com/underwire/2013/08/qq_netflix-algorithm/

http://www.slideshare.net/erikbern/collaborative-filtering-a...

http://benanne.github.io/2014/08/05/spotify-cnns.html https://news.ycombinator.com/item?id=8137264

http://www.playdar.org/ https://github.com/RJ/playdar-core http://news.ycombinator.com/item?id=3876724

https://github.com/bmcfee/librosa

http://forever.fm/ http://blog.petersobot.com/introducing-forever-fm

http://musicmachinery.com/2011/05/14/how-good-is-googles-ins...

http://musicmachinery.com/2012/11/12/the-infinite-jukebox/

Btw., as I am developing my own music player, my main important feature/goal is a kind of automatic DJ which automatically selects music and I also especially had a discover mode in mind. However, that main feature did not evolve that much so far because there were so much other things to implement first. For now, it only supports access to files in your file system, and it looks on tags and artist and adds quite some randomness to it.

https://github.com/albertz/music-player/blob/master/WhatIsAM...

The approach you describe is one of many. Traditional collaborative filtering using matrix factorization will "discover" the features you listed, or other non-intuitive features, from patterns in the data. They don't always need to be listed explicitly, and training individual asymmetric models (like your one NN per person) isn't always ideal. Per person models won't make a lot of sense unless each person has rated a large number of songs, which usually isn't the case.

In general, methods like this require an epic amount of data to work well for consumers. The Netflix Prize made a dataset with 110 million ratings available, and that was barely enough to make impressive predictions. The issue here isn't the statistical model being used, but the paucity of training data. If you trained this or any other model on one billion rated songs, the quality of the recommendations could blow your mind. They would however still skew toward the average musical tastes.

I've long wondered about other psychological favtors that come into play as well. I wonder if each time I give a song a thumbs up, say, during radio play, they record features like day of the week, month of the year, season, the weather that day, time of day, maybe a geographic classifier like city/suburb/wilderness. These are all things that could be used to make localized, time dependent and situational recommendations. I like a lot of the 'mood' playlists Spotify offers as well, but I've used players in the past where you have to choose between very similar and highly subjective things like calm vs tranquil to populate a mood based list, and I wonder if features I mentioned could make this a more passive process on the user's part.

Agreed. I think suggestions should be based on mood and emotion.

For example if I like 'dark' music I might as well like a song from a techno artist and a classical orchestra.

But this can be very complex. What some people find aggressive music others find a little too soft.

EDIT: by the way: because of the above I also think a recommendation for a song is better than for a band.

That sort of describes the Pandora approach. Each song has a set of "genes" that describe its tone, musical instrumentation, key signature, time signature, rhythmic syncopation, etc. these genes are then used to find similar music. I've found it to be quite pleasing to discover new music when it's seeded with a relatively uncommon song that I like.

For instance, I added "Going Through Changes" by "Army of Me", and its recommendations included songs from bands I'd never heard of and actually enjoyed, like Radford, Red (a Christian band, one, as an atheist, I wouldn't ever have browsed on my own, but ended up really enjoying), Maxeen, Black Lab, and Brand New. Each had "genes" that made their algorithm align them with the seed song.

I don't use it much right now - I find myself in a cycle of discovery and then stubborn re-listening of the same four or five albums - but I expect I'll go back to it the next time I'm back in that exploratory mood.

PageRank is analogous to this type of music discovery if you created a graph based on artists who have collaborated together. It'd be about the same as having a teleportation vector that always restarts the random walk back at the artist you're wanting to explore around.

A class I took this fall covered recommendation on attributed graphs. On of the research papers on the reading list for the topic was very interesting: it extends the matrix factorization algorithm on a bipartite graph (of users and soundtracks, for eg.) to more general graph structures (which could include soundtracks, artists). I think this exploits local connectivity between artists much better than the vanilla matrix factorization algorithm.

[1] Yu, X., Ren, X., Sun Y., Gu, Q., Sturt, B., Khandelwal, U., Norick, B., Han, J. (2014) Personalized entity recommendation: A heterogeneous information network approach. in J Proc. 2014 ACM Int. Conf. on Web Search and Data Mining (WSDM’14)

LDA actually works better for this in my experience. That is if assessed by the median distance to leave-one-out from the ones liked by a user in the test set.

Nice, except it is artist-based, not song-based.