Why Netflix Never Implemented the Million Dollar Winning Algorithm
techdirt.com
techdirt.com
Number of words from the OP: 374.
What's worse, the OP is just paraphrasing the original post and not challenging it one bit. I dispute the fact that users' preferences or viewing strategy are different for DVD as for streaming.
Sure, one can "try" more with streaming, and the opportunity cost of trying is a little less for streaming than it is for DVDs -- for DVDs, renting a "bad" movie takes the place of a good one in my cue, whereas trying a streaming show costs me nothing but my time.
But here's the thing, though: the cost of my time stays the same. I hate spending ten or twenty minutes watching something that turns out uninteresting (to me); it upsets me so much that the mood is ruined and often I just go do something else.
I love Netflix for the quality of its recommendation system (unparalleled, in my experience); it's an incredible moment when I watch something I would never have found on my own, that I truly enjoy.
It would be a shame if Netflix stopped regarding this as their most precious asset.
There is no realistic way to translate all the complexity of ML production environment into a neatly packaged problem for contestants. You have to cut down dataset size, drops features, simplify targets and trivialize the implementation requirements (computational complexity, prediction latency, engineering cost). Not to mention that you have to expend significant effort to prepare / normalize / sanitize your dataset and even then you could still get hit with a privacy lawsuit, just like Netflix did.
The result of the contest at best is a set of "ideas" that you have to review/study before starting your own implementation / experimentation from scratch. Did Netflix prize pay for its cost? Perhaps the biggest benefit was generating publicity and helping with data scientist recruiting.