So it’s finally done. I have a ~4GB SQLite database file that has all user-movie-rating triples, along with indices on users and movies. It took almost a week to get all that data in and indexed. Most of that time was spent indexing. My next moves will be creating tables with per-movie and per-user average ratings. After that, I’m going to start working on SVD.
Recently in NetflixPrize Category
Since this is such breaking news at this point. SQLite is really really slow creating indices on big tables. It’s still trying to create the first index on users. I’m about to kill it, wipe the database out, then set the indices up to be created while inputing the data. Maybe that will go faster.
Importing the ratings took about 2 hours. That was with a little tuning, too. Now I'm trying to generate those indices. It's going to take just as long. Ugh. Too much data.
So, I took another look at the training data for this contest. Absolutely freaking enormous. 100 Million ratings from 17000 movies and nearly 500K users. Unfortunately, the user ids run from 1 to 2.6M and have lots of gaps. So, importing the ratings as a matrix in a C app is not an option - the matrix would be over 8.5 GB, even using just chars. So, I thought SQL might be good - the biggest requirements here are being able to index in like a matrix. SQL allows that well enough, especially with indexes on both of those columns. I started with SQLite3, and it seemed slow. So now I'm trying MySQL. It's slower. A lot slower. Either I suck at configuring it (possible, but it shouldn't be this bad out of the box) or it's just slower than Christmas for lots of INSERTs. I may go back to SQLite. Either way, I need an easy read-only DB for this part. CDB seemed like a good option - they're really quick, read-only, and even fast to create. The downside is that is a hash-table only. Since I need a matrix-style system, that seems like a bad plan. I could just use "userid.movieid" as the key, but then I can't read in all the ratings from a user or ratings on a movie. Even duplicating the data in different views doesn't help. This needs to be either SQL or a matrix - there are things I'll need to do that only work with those access methods. The rest of this program can be done in C/C++ without databases, but this really needs to be a DB. Definitely going back to SQLite though. And then I'm going to let this run overnight. Because it's going to take that freaking long to import this data.
So, the Netflix Prize has me intrigued. Really, really intrigued. I'm not good enough at machine learning, stats, or whatever to win, but I'm going to be playing with it in the future. Code to follow. I'll be implementing something like this at first.
