One that I really want baked:

Being a music lover I have always wanted a set of CDs given to me which contain hundreds of songs which I have never heard but will definitely like. Someone should design an algorithm which will extrapolate the types of music you like down to a tee and come out with a number. This will be based on your existing music collection (The more CDs you have, the more accurate the extrapolation. I have about 200 so I should be able to get a decent estimate). Then you could visit a music store and type the number and a computer will display the tracks it thinks you will like most.

For example, if you are a jazz lover the part of the number which refers to the number and length of self-indulgent solos. Status Quo lovers will want the "number of chords" index to be low. I'm sure you can think of more than I can.

This number, incidentally, will probably be pretty long so you might need it on disk so you don't get long queues at the machine.
-- afroman, Nov 01 2001

Google search for Firefly http://www.google.c...&btnG=Google+Search
Web debris left behind from the now-defunct Firefly. [hippo, Nov 01 2001, last modified Oct 04 2004]

The cut-throat world of music recommendation http://www.newmedia...les/NM01040178.html
[hippo, Nov 01 2001, last modified Oct 04 2004]

Music Retrieval by Similarity http://www.fxpal.co...te/musicr/doc0.html
Rmutt, in his spare time, is working on a system that analyzes the audio (no collaborative filtering). Rmutt is not promising anything -- your mileage may vary. [rmutt, Nov 01 2001, last modified Oct 04 2004]

emusic and RealPlayer partnership http://emusic.com/
Recommendations Are 100% Accurate based on downloads I've done of Blues and Jazz [thumbwax, Nov 01 2001, last modified Oct 04 2004]

Finding similarities based on semantic analysis of lyrics http://74.125.155.1...+&hl=en&as_sdt=2000
Doesn't work very well apparently [fatgriller, Nov 25 2009]

Time Warner and/or Amazon.com would love to hear from you (of course, you have to have bought your 200-odd CDs from them in the first place, which might possibly be your point). As far as books go, Amazon starting getting my predilections right after only a few purchases, so try them until something better comes along.
-- DrBob, Nov 01 2001


There used to be a very good CD recommendation service on the web called 'Firefly' - I think they folded and their database got bought by Amazon. Basically, you entered in lots of CDs you owned, as did lots of other people. Based on various "Most people who own X also own Y" algorithms, it would then recommend music you should try. I entered in about 30 CDs and it then recommended loads of other stuff which I already had and liked, so it was pretty accurate. Amazon may use the Firefly software now.
-- hippo, Nov 01 2001


What the astute [afroman] is suggesting is much, much cooler than lame data mining like Firefly et al, and happens to be something I began research into this past summer. It's similar problem to speech recognition & surprisingly totally doable* - I've been getting about 90-95% on my testing data (culled from mp3.com), and can probably get that to 95+ w/ boosting. As long as you increase the mean quality of what you eventually review, you win.

*The downsides are that it obviously can't handle interpretation of lyrics e.g. Radiohead or TMBG (which might just be mediocre audio otherwise), and if you want to detect novelty, it's easy on the face, but actually kind of a hard problem - you don't actually want white noise unless you're a fan of Richard D. James' more recent works.

Misc details in case anyone's interested: it needs about a dozen training songs (you don't have to point it to "number of chords"; it Just Knows as long as your dataset is representative), trains on random 20 second snippets. Training is a little slow; ~5 min per model, but is a one-time deal unless you want to change it. Testing a song is fast after that ~15 seconds. If anyone is actually interested in pre-pre-alpha testing something like this, reply to this annotation.
-- prometheus, Nov 02 2001


How do you access the sort of repository of potentially likeable songs you'd need in order to test them; wouldn't someone need to have built some sort of online database, or summat? I'm intrigued.
-- Guy Fox, Nov 02 2001


[waugs] Aha a challenge - I promise you it's very real. I'm a grad student in Cognitive Science with a focus on machine learning, which granted is just an appeal to authority (and not a very good one), so I invite you to test its performance for yourself.

all: email me at the address in my profile and I'll hook you up.

I've set up a small website that hooks into the "MP3.com Music IOS API" - you rate songs from your desired genre. I have the training stuff online and will be working on putting a web interface on my testing (as in "post-training classification") code this weekend. Rate songs from your genre for my training set, and hold some back for a blind test set. You wouldn't have to listen to each entire song; just enough to classify it as good/bad.

[guy fox] Yeah, you have to manually classify the original few dozen training songs, but the problem with the other systems is you have to have another human besides you in the loop so you can't automatically classify completely new songs, & as was noted, performance is crap anyway.

I'm actually slavering at the prospect of more data - until now only person besides me has shown any interest.
-- prometheus, Nov 02 2001, last modified Nov 03 2001


This service already exists (in a way): www.audioscrobbler.com (although it is currently being upgraded). It's designed for people who listen to loads of mp3s on their PC's rather than you entering a list of CD's. All you do is register on the site, download the plugin for your relavent meadia player and it monitors what music you listen to and how often so you don't even have to think about writing a list of what you like. It then sends this information back over the internet (if you aren't connected at that time it stores the data in a cache). Once you have listened to several hundrend tracks it will give you a similarity list which tells you who else likes the same music as you and suggests new tracks that they like that you also might like. Obviously it is slightly more complicated than that but you get the idea. I hope I've explained it well enough; you can go to the official site and get a full explanation when it's up and running properly.
-- TsunamiCD, Oct 25 2003



random, halfbakery