Check that ALS convergence is strictly decreasing

[kafkasl]

ALS convergence when evaluating the RMSE of the training dataset should decreasing.

In some executions (namely, netflix dataset with 20 factors and lambda 0.065) it appears that the RMSE increases for 4/5 iterations before starting to decrease.

The tests of ALS show strictly decreasing RMSEs.

This shouldn't be possible in the original ALS. However, this could stem from the implementation details such as:

• Parallelization
• This version uses "scores" for each movie, some versions use boolean (seen vs. not seen).
• To update the chunks the conjugate gradient method is used which I believe is an approximation