There's a machine learning competition running here in Salt Lake City that deals with predicting the popularity of Taylor Swift videos from snippets of music. And by snippets I mean 1/10th of a second of a song. The goal is to classify whether a song is a 'success' or 'not' (if you can take 30m views as not being successful and 300m as success.)
In marketing we try to understand what makes a consumer tick, but in this case an audio track has been converted to a series of numbers that look something like the following. Now imagine 3,300 columns of numbers for each sample...
Success V1 V2 V3 V4 V5 V6 V7 V8 V9
0 -5787 9566 -511 -2274 18589 1170 2232 5073 -2578
1 1067 - 521 1524 -209 -957 -777 -2666 716 -3273
1 397 -314 -1701 -2568 -463 1123 1041 916 789
1 784 -1017 2038 4134 1453 3731 3644 2759 -2033
0 -19632 -13344 -14428 -13195 -3940 -1306 -1206 -1303 -1256
This is where you have to trust the machine and the team. There will be no conference room debate as to what impacts success - the math just works. And that raises an important question: how do we now improve the consumer experience and deliver on the brand promise?
Keeping with the music theme, a musicologist can deconstruct the signals into themes and styles much in the same way that Pandora was created from a music genome. As a result the experts won't be just the ones who can apply a Restricted-Boltzmann-Machine to a classification problem, but those who can translate the results. Interpreters will be in as much demand as the coders.
Keeping my Rosetta Stone polished.