'Danceability' algorithms are flawed

RobG's Avatar

RobG

27 May, 2020 07:07 PM

Basing Danceability on DFA is like asking AI to write Christmas carols.

This + This currently equals...

  1. 1 Posted by RobG on 28 May, 2020 06:24 PM

    RobG's Avatar

    More from the algorithms...

    The detrended fluctuation analysis (DFA) has been proposed by Peng et al. and was first applied in biomedical analysis. Jennings et al. introduced the method to the field of music analysis by using the DFA exponent for musical genre classification. In this paper we further exploit the relation of this low-level feature to semantic music descriptions. The feature has been computed on a large-scale collection of 7750 tracks for which manually annotated semantic labels like "energetic" or "melancholic" where available. Associations with high statistical significance could be found between some of these labels and the DFA exponent. The findings sustain the hypothesis that this feature can be linked to a musical attribute which might be described as "danceability". source

    So Herrera and Streich used a "large-[read:laughable]-scale collection of 7750 tracks" and it gives us this (see attached image).

    Here are the two tracks, both with a score of 3 low:

    As a user, I want to get more accurate results from analysing my library with beaTunes.
    As a data-nerd and music lover, I want to clip Herrera et al round the earhole and ask WTF???

  2. Support Staff 2 Posted by hendrik on 29 May, 2020 09:57 AM

    hendrik's Avatar

    Hey Rob,

    The danceability algorithm by Herrera et al. is certainly not perfect, but interestingly there hasn't been much more recent work. In 2005, when Herrera et al. published their algorithm, 7,750 annotated tracks were actually a sizable collection. When you think of tracks as many potentially overlapping 10s snippets (you should not need more than 10s to determine whether a song is danceable), you may even argue that 7,750 tracks equals many more training samples for a machine learning algorithm, which might not be so bad after all.

    That all said, yes, you are right. This is not the best solution and it needs work. It's on the Todo list for beaTunes 6.

    -hendrik

  3. 3 Posted by RobG on 29 May, 2020 10:29 AM

    RobG's Avatar

    That'll be some task, I bet! The key issue, of course, is our unique tastes as humans. Some classic shoegaze indie will be 'danceable' to some, but probably not to the majority of pop music consumers.

    This issue that's arisen from the lack of continued research reminds me of a situation about 20 years ago when I was DJing at a pub with some friends. It was mid December and an office party was underway. A punter rolled up to my mate and screeched "Can you play something a bit more funky?"... He was playing Funkadelic at the time! I believe his reply was "Sweetheart, music doesn't get anymore funky than George Clinton!"

    As someone who plays 'open genre', I know that I'll never be able to rely on this, nor any future algorithm. If I point the finger at tech, it's just too subjective and emotive for AI. If I'm honest with myself, I like music that the general public can't stand unless it's buried in the soundtrack of Stranger Things! LOL

    Knowing that I'll be dependent on manually scoring the majority of my music, is it possible to apply the Danceability score by keyboard shortcuts (similar to the Alt+1-5 rating from the Edit menu)?

  4. Support Staff 4 Posted by hendrik on 29 May, 2020 10:37 AM

    hendrik's Avatar

    Yep, there is definitely some subjectivity involved. A field like that it is nevertheless useful for manual annotation or getting some hint from an algorithm, that should be taken with a grain of salt.

    BTW: If your files have been processed by MiK, it may have written an "energy level" into the files. beaTunes picks up on this (on initial import) and displays it as "danceability".

    Knowing that I'll be dependent on manually scoring the majority of my music, is it possible to apply the Danceability score by keyboard shortcuts (similar to the Alt+1-5 rating from the Edit menu)?

    No, I'm sorry, there is no pre-defined keyboard shortcut for that.

  5. 5 Posted by RobG on 29 May, 2020 09:08 PM

    RobG's Avatar

    Oooo! Good to know the impact of MiK. Sounds positive. The proof will be in the pudding, I guess. I didn't plan to process with MiK for a while, but knowing this I may test a small selection of files. Maybe 7,750 or so? LOL

    No, I'm sorry, there is no pre-defined keyboard shortcut for that.

    Possible beaTunes 6 feature request. I'll see if MiK improves the danceability scores

    Thanks for all your help!

Reply to this discussion

Internal reply

Formatting help / Preview (switch to plain text) No formatting (switch to Markdown)

Attaching KB article:

»

Already uploaded files

  • DFA_is_DAF.jpg 31.4 KB

Attached Files

You can attach files up to 10MB

If you don't have an account yet, we need to confirm you're human and not a machine trying to post spam.

Keyboard shortcuts

Generic

? Show this help
ESC Blurs the current field

Comment Form

r Focus the comment reply box
^ + ↩ Submit the comment

You can use Command ⌘ instead of Control ^ on Mac