Building a recommender system is a lot harder than just filling in a matrix with missing values. One has to deal with the constraints of modern web. People don't want to spend time and don't want to think. The recommendations don't only need to be relevant, but also divers and robust. I touch upon those hard to measure aspects with lessons learned from my recommender system www.namesilike.com
5. Wout Seppe Gust
Senne Ward Nand
Tuur Ferre Siebe
So if a user clicked on:
We can recommend:
Korneel Staf Kamiel
1. What do these names have in common?
8. Caroline Céline Stéphanie
Virginie Sandrine Myriam
Sonia Christelle Claire
So if a user clicked on:
We can recommend:
Valérie Aline Karine
2. What do these names have in common?
9. 3. What do these names have in common?
Oscar Sem Gust
Ilyas Milo Kamiel
Owen Miel Remi
11. So if a user clicked on:
We can recommend:
Muhammed Luka Ibrahim
3. What do these names have in common?
Oscar Sem Gust
Ilyas Milo Kamiel
Owen Miel Remi
14. Recommendations need to be
Relevant
Diverse
Serendipity
Robust
i.e. make the user feel like you understood him
15. Solution
Early stage: Business rule
If not clicked on a arabic name, do not predict one
Later: Ensemble model
Combine time series features with latent features
37. Feedback loop?
1. Start with the ambition to build a general bookshop
2. Your recommender system recommends mostly cook books
3. People who love cookbooks love your recommender system, they
invite other friends
4. Your measurement of satisfaction of your recommender system
starts to be biased because your users are mostly cook book
lovers
5. You end up with a cook book shop
39. There is nothing important or of great scientific
merit in this problem, they will sneer. These great
minds should have been putting their immense
brain power to something that would benefit
mankind.
Unfortunately, whatever non-scientists may think
about science and its importance, my experience
has been that most (?data-)scientists do their
research because they are interested in the
results and get intellectual excitement out of the
work. Seldom do good scientists think about the
eventual importance of their work.