Machine learning can solve 80% of the problem, but it's still important to pose it well. What Netflix needed wasn't just better predictions, it was better retention.

Algorithms alone are a poor competitive edge. With data and a better problem definition, they are a huge competitive edge.

Or maybe their success is due more to smart engineering than research optimization.

WAY wrong directions.

‘Machine learning’ is a junk field because it has no solid rational foundation and no powerful methodology. About all the field is is heuristics and bad applications of cookbook statistics 101.

I worked in the field at the Watson lab in Yorktown Heights and each day had to hold my nose not to upchuck from the stench. Finally I took our central problem, found a solid solution, and published it.

E.g., ‘machine learning’ keeps looking for an ‘algorithm’. They are already lost, digging in the wrong place. By analogy they would look for an ‘algorithm’ to say how to navigate a space craft from Earth to a selected spot on a selected moon of Jupiter. Laughable. Instead, start with Newton’s law of gravitation and laws of motion and some ordinary differential equations. For the software and computing, that’s just to do the arithmetic. There’s no ‘algorithm’ (unless want to count some numerical techniques for solving the differential equations). What the computer science people are missing in ‘machine learning’ is analogous to Newton’s laws.

It’s possible to do much better on the problems being attacked by machine learning, but the computer science community doesn’t know how to proceed. The needed techniques are rock solid but they are quite advanced. Nearly no one, even at the top of research computer science, has the prerequisites because they didn’t take the right courses in grad school. The fields that understand the needed techniques believe that as research ‘machine learning’ problems are trivial and that too much is already known.

Interestingly enough, it was the mistakes the algorithms made that lead to the most significant discovery. I took all the misclassifications and examined them myself in an Excel spreadsheet. In doing so, I found an intriguing pattern which ended up adding a lot of value to our study.

I humbly submit Netrics as an example of freshly applying Machine Learning to deliver value and create market advantage.

Also, using the Netflix Challenge as a single data point to flog all of ML is not exactly fair.

The impact and importance of ML will only grow – these are a very exciting times!

–Ian

More here:

http://thenoisychannel.com/2008/11/21/the-napoleon-dynamite-problem/

If you have any examples of areas where ML was freshly applied to create an advantage I’d be really interested to hear about them.

Especially less mathematical areas people tend to stay away from advanced methods due to a lack of understanding (on both sides, application side does not know about the powers of ML and the ML community has no clue what the application side needs).