"Liking is based on deep mathematical patterns"

In the video Music Intelligence: 2012, Malcolm Gladwell (Tipping Point, Blink), interviews Mike McCready, CEO of Platinum Blue.  Platinum Blue is a company that has a "hit predictor" as well as a hybrid music recommender (using social and 'mathematical similarity').    

I must admit that I've always been quite skeptical about hit predictors, and after watching this video I am even more skeptical.   Here's an example that set my bullsh*t detector off.

First we see a visualization of a music space:


 

Then, Mike shows us the same music space with all of the non-hits removed,  this leaves a plot that shows just the hit songs:

 

It is rather amazing isn't it how well the hits cluster -  and the clusters have such nice sharp boundaries too.  It's almost too good to be true - no wait, it *is* too good to be true.

Platinum Blue claims that their hit predictor is 80% to 85% accurate.  That's nothing, I can build a hit predictor that is 99.9% accurate.  No matter what song you give my hit predictor it says "not a hit".  If you run the 500,000 songs released last year through my hit predictor, it will only get about 500 of them wrong (the ones that actually became hits).

Listening to this interview makes my skin crawl - I just don't buy it, and I'm rather disappointed with Malcom Gladwell just accepting this without showing any skepticism.

Posted on: Jun 01, 2007

Posted by: plamere

Category: music

Permanent link to this entry

Comments:

Wow, my brain must be a computer, because when I heard Norah Jones' "Come Away With Me" I thought it would be a hit too. Man oh man. Discover a Gnarles Barkley or a Strokes or a Modest Mouse or an "Oh Brother Where Art Thou"...something unexpected and out of the ordinary.

Being proud of your software for predicting that Norah Jones album would be a hit is like being proud of looking out the window on a hot sunny day and predicting that the pool will be busy.

Any of these companies (another one is Music Intelligence Solutions) who claim to have come up with a system to mathematically generate hits has never been to a sweaty bar at 1:30 in the morning covered in somebody else's cigarette smoke just basking in the raw power of the music that is pouring out of the speakers.

Bleh.

Posted by Zac on June 01, 2007 at 08:37 AM EDT #

Everybody here needs a ROC analysis class and the folks at platinum blue as well...

Posted by DrMusic on June 01, 2007 at 12:03 PM EDT #

Maybe, by 85% accuracy, he meant 85% mean average precision. Or precision at 100% recall. The latter makes more sense, actually. Suppose, out of 100,000 songs, there are 100 hits. By the time you've extracted 100th hit, how many non-hits have you also extracted? If the answer is somewhere around 18, then you have 85% accuracy. Natch? I think you're thinking of this too much as a classification problem, where you're interpreting "accuracy" as a correct label (hit vs. non-hit). If, on the other hand, you take the information retrieval perspective, and think of it as the problem of selecting a set of hits from a large collection of non-hits, then all that counts in that score is total-relevant divided by total-retrieved. Or, total-true-hits-extracted divided by total-extracted. D'ya see what I mean?

Posted by 72.255.12.185 on June 02, 2007 at 12:02 AM EDT #

Yeah, I see what you mean. I understand too that when giving a talk to the general public it is best not to talk about ROC, F measures and all that, since 85% of the people won't know what you mean. My point is that saying you have an 85% accuracy is meaningless for a hit predictor.

Looking at the Platinum blue FAQ we see this question and answer:

Q: What is the accuracy of Music Xray™?

A: Most singles released by music labels sound and feel like hits but lack the optimal mathematical patterns. This is a main reason why the industry has less than a 20% success rate. Less than one in 5 songs that are released as singles and promoted iwth(sic) a significant budget actually reach the charts. By using Platinum Blue’s technology, hit rates can be increased to over 80% - four times the current industry rate - effectively providing a risk management capability that helps yeild(sic) significant improvement in return on investment.

But this doesn't tell us much either. The only thing I learn is that my song may lack the optimal mathematical patterns.

Posted by Paul on June 02, 2007 at 08:11 AM EDT #

the poor a+r managers -do they still exist? - have been so scared in the past that some of them bought this polyphonic hmi stuff, ... and they had never heard anything about ROC ...

Posted by Stephan on June 04, 2007 at 08:47 AM EDT #

Post a Comment:
Comments are closed for this entry.

This blog copyright 2010 by plamere