Saturday, June 27, 2015

Machine and learning, trial and error

His Master's Voice (HMV)
In the winter at Strata West I had a chance to see Oscar Celma of Pandora discuss machine learning from the perspective of the most established streaming music company. Because at root Pandora has a lot of human intelligence about music, its machine learning applications of song suggestion are even more interesting.

 The thing I picked up from Celma's presentation was that you can only get so far with your basic breed of suggestion engine. In the radio days a big voice intoned 'don’t touch that dial'. Now something else is in order.

 You see, if you play the straight and narrow and give them what you know they want, they get bored, and tune out. The element of surprise has been intrinsic to good showmanship immemorial . The machines can get better and better, but at a slower and slower rate. People eventually want to come across a crazed Jack Black pushing the 13th Floor Elevator button in HiFidelity.

The machines have trouble contemplating the likelihood that a viewer may be prone to enjoy Napoleon Dynamite, as was précised in this story about the 2008 NYTimes Magazine story about the Netflix algorithm contest that fate (my brother cleaning the upstairs) cast upon my stoop.

 TO BE CONTINuED
Posted by at No comments:

Sunday, June 21, 2015

Momentous tweets for a week in June 2015




Posted by at No comments:

Sunday, June 14, 2015

Advance and quandry: Big Data and veteran's health

The era of big data continues to present big quandaries. A Time's story, Database May Help Identify Veterans on the Edge, covers the latest brain teaser.

The story points to new research, published in the American Journal of Public Health, researchers at the Department of Veterans Affairs and the National Institutes of Health described a database they have created to identify veterans with a high likelihood of suicide, as the Times story (Fri, Jun 12, 2015, p A17) points out, " in much the same way consumer data is used to predict shopping habits."

The researchers set up a half a database that comprised variables associated somewhat with suicide cases between 2008 and 2011. They ran what I assume to be a machine learning algorithm on that. They then tried to predict what would happen with the remaining half of the database population. They then concluded that predictive modeling can identify high risk patients no identified on clinical grounds.

But predicting suicide is not like predicting likelihood one might buy a Metallica song, is it? How does the doctor sell the prognosis? "A machine told us you are likely to commit suicide."? Certainly some more delicate alternatives will evolve. A lot of the variables – prior suicide attempts, drug abuse – seem patent. Maybe doctors have just been more likely to guess on the side of life. If the Chinese government hacks the database, and sells the data, will the chance of suicide follow you like an albatross, and fulfill itself ?

Like so much in the big data game, the advance carries a quandary on its shoulder.

Related
http://www.nytimes.com/2015/06/12/us/database-may-help-identify-veterans-likely-to-commit-suicide.html
http://ajph.aphapublications.org/doi/pdf/10.2105/AJPH.2015.302737
http://ajph.aphapublications.org/doi/abs/10.2105/AJPH.2015.302737

Posted by at No comments:

Saturday, June 13, 2015

New and notable Week of Jun 8







Posted by at No comments:
Subscribe to: Posts (Atom)