A Modern Recommender Model Architecture

20 points by Ameo

jyc

Very cool!! Thanks for writing this up. I was working on my own much much dumber content-based (as opposed to collaborative-filtering) recommender system a while back just using regression on embeddings.

One thing I noticed is that when I plugged in some anime I liked, the suggestions all seemed like very popular anime. I’d actually heard of all of the results and either already watched them or previously decided against watching them. When I think about it this makes sense from a Bayesian POV but it would be cool to be recommended surprising/lesser known anime instead.

Ameo

Thanks for checking it out!

One thing I noticed is that when I plugged in some anime I liked, the suggestions all seemed like very popular anime.

I see. There may be some cases where this new model is still tending to do this, maybe particularly in cases with small input profile size like the interactive recommender. If you are interested in posting or sending me what shows you put in, I'll take a look and see if I can learn anything deeper from the scoring going on under the hood. No worries if not though.

Also, in the non-interactive version of the recommender that takes MAL profiles as input, there's an "advanced options" section with options that give some manual control over the scoring and popularity weighting as well. I didn't expose it for the interactive recommender because I was finding that most users were on mobile and many seemed to be younger as well, so anything more than the simplest possible UIs tended to confuse then and cause them to abandon the tool.