4 Comments
User's avatar
Vishal's avatar

Nice article Gaurav! Section 1 says that we can read more about graph based recommendation systems in the "Further Reading" section. Can you please add one.

Expand full comment
Andrei Lopatenko's avatar

Would be interested to see details of efficient retrieval systems for two-tower networks

Expand full comment
Gaurav Chakravorty's avatar

Thanks for the question.

Both in the two tower model and in GNNs, for efficient retrieval, one can efficient maximum-inner-product-search systems like ScaNN ( https://github.com/google-research/google-research/tree/master/scann ).

I have a COLAB here that shows how to use ScaNN : https://colab.research.google.com/drive/1nGweHc8_ytTgQrI8g0oRgjgk0o6vIJJt#scrollTo=euffs3pj8X4O

Also please check out an implementation by Pinecone for efficient nearest neighbor search : https://www.pinecone.io/product/

Expand full comment
Andrei Lopatenko's avatar

Thank you

Expand full comment