Use ALS instead of SVD? about elasticsearch-spark-recommender

[MLnick]

Hi there

Indeed, SVD and ALS are very similar in concept - they are both factorizing the ratings matrix into sub-matrices. It's debatable which is best and it really depends on the situation. In fact for most real-world problems one should try a few algorithms and approaches to see which one best fits the objectives.

In this case, the ALS implementation in Spark is performant and scalable and is the easiest to use and get started with. The SVD implementation available in the older RDD-based Spark APIs is not very scalable by contrast. Hence the reason we use ALS here.