Collaborative Filtering and Alternating Least Squares

This would be fun / maybe useful to people.

ALS has a couple embarrassingly parallel stages. But getting good performance in a distributed setting requires some very careful bookkeeping about which blocks of the user and item factor matrices are required at each step. https://github.com/dask/dask-ml/compare/master...TomAugspurger:als does none of this bookkeeping at the moment, and is extremely slow.

I haven't found any papers about blocked ALS yet. The Spark scaladoc gives some hints: https://spark.apache.org/docs/2.2.0/api/scala/index.html#org.apache.spark.ml.recommendation.ALS

Some references

- Hu et al 2008: http://yifanhu.net/PUB/cf.pdf
- Winlaw et al 2015: https://arxiv.org/pdf/1508.03110.pdf
- https://github.com/benfred/implicit

I plan to work on this for fun on nights and weekends, but if someone wants to take it up then by all means.