[prades]

I've built a recommender system for my movie database app "Coollector Movie Database". It's based on Collaborative filtering and it took me 2 years to implement. I built it from scratch and it's unique in several ways (for example, you can view the reliability of each recommendation). The technical difficulty is to crunch fast enough a huge quantity of data. I've had to apply all the optimizations that I could think of.

https://www.coollector.com/help.html#recommendations

[superasn]

That is a pretty amazing site! Can't download anything right now which makes me think It would make it a lot easier if instead of a downloadable software it was being offered as a saas site.

[prades]

Thank you! It could not be free if it was a website because calculating the recommendations uses quite a lot of CPU. It's fine when run on each user's computer, but it would require expensive servers to make all the calculations for everyone, with a risk of congestion.

There's a website named criticker.com which gives great movie recommendations, but you'll see that they have problems handling all the calculations. You'll literally see the recommendations being slowly generated, and updates are a problem when you rate more movies.

[eggie5]

How would you generalize the method you are using?

[prades]

I don't like ML frameworks (Tensorflow, etc...), maybe it's because I haven't tried them. My understanding is that they're like a magic black box: you input some data, you adjust some settings, and you wish for the results to be good. Instead, I've taken a direct approach to the collaborative filtering problem, the difficulty being to correlate a huge amount of data. Some said that only quantum computers would one day be fast enough to solve the recommendation problem, until recently a student demonstrated that it could be solved with classical computers.

https://www.quantamagazine.org/teenager-finds-classical-alte...

This student's algorithm is quite different from mine, but I suppose that my algorithm is yet another example of solving the recommendation problem with classical computers.

[prades]

As long as the users rate how much they like something, my method could easily work with songs or books or anything.