Today I went to give a talk about our library to colleagues from
other departments working in another office. There are quite a few
systems which might masquerade as recommenders not being ones.
Everybody knows that Amazon employs recommender systems for their
goods. But if you sell only twenty or fifty products do you need a
recommender? A classifier will do a better job for sure. You also
don't need a recommender for choosing an item with which a user will
interact for the second time. Recommendations are usually about
something new in an ocean of options. Choosing what track to replay
from your history is not a task of constructing a playlist given
millions of songs. Again, for this problem, a classifier can provide
a plausible solution. Even item-to-item recommendations are fading
away as a subset of metric learning. In a 'customers who bought this
also was impressed by that' task, the deep learning revolution has
happened already. That is not so in personal recommendations where
you still need to accommodate millions of users with shortlists of
items chosen from millions of items, and preferably do that in
real-time.
