Slashdot Mirror
Netflix Announces Second Data Mining Contest
John Snodgrass writes "Neil Hunt, Chief Product Officer at Netflix, has announced on the Netflix Prize Forums that they are planning to hold a new data mining competition. The second competition will have some twists and is expected to be shorter in duration. It will feature two grand prizes, to be awarded in a 6 and 18 month time frame. A previous competitor still active on the board has already dubbed it: 'The Sparse Matrix: Reordered' and 'The Sparse Matrix: Factorizations.'"
> Why do tens or possibly hundreds of thousands of dollars worth of work just for the chance that you might get payed? It seems absurd.
Challenge and notoriety.
For that matter, just about everything you do has a chance of failure, so why do anything?
=Smidge=
Most of the time people want to tinker/play with concepts. They just need some sort of motivation to get them going. The chance that you might get payed is apparently enough for some people.
"When life gives you lemons, don't make lemonade. Make life take the lemons back!" -- Cave Johnson
Most of these are research groups that would publish their results and research anyway. This gives them a practical application and a chance for some fame and money -- the research still gets done and published.
I'll add one more thing. Netflix has done the community a favor by providing a large dataset for testing algorithms. Data mining requires data. It requires more than just raw data. It is really difficult to know how well your algorithm works without data that has known answers to compare to. A good test dataset lets you compare your results to other results.
I used to think I was unique in what rare movies and music I liked until I met someone who had almost an identical collection to me. On top of that, we both had some of the same clothes. The reason netflix researches these data mining techniques is because our tastes really do cluster into groups. For some it might be because they like DeNiro films and Spaghetti Westerns. For others it might be that they like two screen writers - though they never know it. The payoff for getting this right if you are Netflix is that if a customer gets recommendations and he/she use those to fill your queue, then they are much less likely to cancel netflix anytime soon.
It allows the researchers to "cheat" a bit too via an argument by authority, which is not always good, but does at least make the researcher's job easier. A big issue in data mining is that it isn't purely a technical field, but one with both conceptual and technical issues. The over-arching goal is something like, "get useful and/or interesting information out of data". But what is "useful", what is "interesting", and how do we measure when we've gotten it or not? Usually you have to defend why your problem is the right one, why your metric is the right way to measure success on it, etc. Working on the Netflix competition lets you sidestep all that, because Netflix has already decreed exactly what the goal is, and what performance metric will be used to judge success at that goal, leaving only the technical problems.
10 PRINT CHR$(205.5+RND(1)); : GOTO 10
The purpose of this contest is to figure out who won the previous contest.