search

maciejkula / spotlight

Deep recommender models using PyTorch.
MIT License
2.99k stars 428 forks source link

LightFM MetaData Embedding layer implementation in Pytorch #144

Open sgaseretto opened 5 years ago

sgaseretto commented 5 years ago

Has anyone done something like the metadata embeddings used in LightFM and make it work? I understand in a very high level way how they are created in LightFM. How ones we have the item features embeddings for each item/user, they are added to create a "compused representation" and then used to perform the matrix multiplication. but I'm having some trouble thinking around how this can be implemented using the Pytorch embedding layers.

maciejkula commented 5 years ago

I had a branch that implemented this: https://github.com/maciejkula/spotlight/tree/hybrid_models

Feel free to mine it for inspiration; I hadn't been able to find an API that felt clean enough.

treper commented 5 years ago

I had a branch that implemented this: https://github.com/maciejkula/spotlight/tree/hybrid_models

Feel free to mine it for inspiration; I hadn't been able to find an API that felt clean enough.

Does this implementation has equivalent performance to lightfm?

aptlin commented 3 years ago

Would love to see any benchmarks if anyone tries the branch out, given that the GPU support should give a little boost in comparison to LightFM.

lgpreston75 commented 3 years ago

@aptlin haven't tested the branch out but: https://github.com/cemoody/simple_mf/blob/master/notebooks/01%20MF%20model.ipynb might be of interest