Netflix prize random whole numbers
General, Netflix October 11th, 2006
Discussing last night’s efforts with Andrew on the way in this morning, we were hashing out why the random algorithms always generated the same RMS values, and the fact that predictions can have decimal values while recommendations are whole integers. I realized I never tested random whole numbers against the recommendations, so I did a quick test. For whole values 1 to 5 on the probe set, I got a RMSE of 1.929. That’s better than my 0.00-5.00 attempt (2.176) but not as good as my 1.00-5.00 attempt (1.748). And that’s about what I expected.
I might need to do some poking around ACM for papers on predictive algorithms. Would Bayes help me here?
Other interesting data I have to play with is the date the recommendations were made, the title and year of the movie, and the general overall numbers for how many recommendations were made for a certain movie. I’m sure I can pull something probative out of that.
About
