# [2021.05.21] A course on building recommender systems

I started writing some code for examples to a planned course on
recommender systems. There are several different packages positioned
as frameworks for recommender systems creation. Usually, they are
highly incompatible with other packages. Sometimes they are not
documented and supported enough (notably ``polara``) or abandoned
(e.g. ``surprise``). Finally, I decided to use the global time split
validation scheme. Recently, it received even more support from the
research community. For the course, I don't need several options to
validate most of the time. For computing metrics, I will rely on my
ex-colleague package, ``rs_metrics``. Even if it's not that
efficient, no one needs such packages for production and big
datasets.