Friday, August 6, 2010

Platinum Blue

I read an article in the New Yorker some time ago, called "The Formula" by Malcolm Gladwell (the author of Blink, The Tipping Point, and Outliers fame) which details a small company providing an interesting service to music labels. This company, Platinum Blue, decomposes a song into its constituent parts such as “melody, harmony, beat, tempo, rhythm, octave, pitch, chord progression, cadence, sonic brilliance, [and] frequency” which are then fed into a computer for analysis. These parts are combined into a song fingerprint, which is then used to facilitate the comparison of these songs. For example, songs using minor keys would be grouped together while songs using major keys would be grouped together, as would fast songs and slow songs.

Hit Prediction

After feeding several million songs into their song fingerprint database, Platinum Blue noticed that hits fell into one of 60 distinct clusters. For example, all upbeat pop songs sung by men might form one group, while Classical would form another (I admit, this is an oversimplification). After identifying these hit clusters, Platinum Blue can tell a record label which of their songs should be released as a single, based on whether the song fingerprint falls into a hit cluster, and is therefore more likely to become a hit. If a song falls slightly outside of a hit cluster, they can also recommend changes to that song so that it would become a hit. Those songs which are far from a given cluster, will never become a hit, and therefore represent inferior opportunities for promotion to their artists and record labels.

Music Xray and Purchase Recommendations

During a different video interview, Platinum Blue heavily emphasized the application of this technology to recommendations for online music retailers… even going so far as to say that they’d waive the licensing fee if it did not increase retailer profitability. While I have not found any indications that Platinum Blue was used in this capacity, the functionality was recently rebranded Music Xray… now facilitating the automated analysis of unsigned songs with the opportunity for submission to major labels. As intriguing as these alternative business models are, I take these developments as a hint that major record labels have not expressed much interest.

Value Quantification

The CEO, Mike McCready, confirmed in a video interview with Malcolm Gladwell (Music Intelligence: 2012) that 80% of all hit songs at any time fall within one of the hit clusters which they have identified. The remainder can be explained by social causes, i.e. the title song to a blockbuster movie, “Eminem’s next release when he was at the top of his career”, etc. Furthermore, “historically, fewer than twenty per cent of the songs picked as hits by music executives have fulfilled those expectations.” McCready adds, “It can cost a million dollars to promote a single in the U.S. market, but nine out of ten of those don’t get a return on the investment, they don’t chart, and executives at the labels are left scratching their heads and wondering what they did wrong.” These statements could be interpreted financially as offering a 10 fold improvement in return on marketing investments, and likely a ten fold return on the fixed assets already associated with the labels, such as employee salaries, offices, equipment, etc. Speaking from a strictly mathematical standpoint, being able to accurately sift through social data to identify 80% of hit songs while reducing failure by 90% is exceptional.


Given that a record label has not bought Platinum Blue for internal use (to thereby improve their profitability), I am curious as to whether the service delivers as expected.

Malcolm Gladwell also makes the interesting point that, if hits exhibit quantifiable similarities in musical structure, then there is skill involved in their creation. This discredits the idea that the music industry tells consumers what to like.

Platinum Blue has the ability to provide music recommendations free from social influences or subject to them. For example, Aerosmith’s song Cryin’ uses the chord progression from Pachelbel’s Canon… so from an objective standpoint, those who enjoy Cryin’ may like Pachelbel’s Canon, if they were not influenced by the social influences associated with the different genre. Could the ability to make such recommendations be used as a test for how rigidly we adhere to our social persona?

I’m also very curious as to how they decomposed chord progression, melody, or harmony into a single attribute which could be compared between songs. What fleeting knowledge of music I possess, tells me that reducing a melody into a single metric which could be compared between songs of different length, tempo, and complexity would be rather difficult… unless there is a “universe” of melodies nested within the song “universe” as well as a “universe” of chord progressions nested within them both?

The CEO said that his methodology is more advanced than recommendations given by "collaborative filtering… which is a bit antiquated at this point. Because you’re only tying yourself into your demographic." I would like to examine that idea more deeply, but that will have to wait for a future post...

Competitive Advantage Commentary

In this instance, quantitative methods were used to inject objectivity into a process, to avoid failures and their associated advertising expenses, thereby providing a 10-fold increase in return on marketing investment. Given the overwhelming appeal of such returns, it is suspicious that Platinum Blue wasn't promptly bought out. While it wasn’t discussed in the article or video, this tool could also be used to reduce signing risk. For example, if an artist already has a number of songs which Platinum Blue identifies as potential hits, then that artist presents less of a signing risk than artists who have yet to prove themselves.

Subsequent blog entries to be made on the other companies mentioned in this article…