dream-faster / laplace-gnn-recommendation Goto Github PK

• graph size: 100,000
• #75
• #74

[] Read https://arxiv.org/pdf/2202.09177.pdf

https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/308919

Some graphs won't be able to handle this, but our encoder-decoder model would be suitable.

eg.: we could fragment out some features like article's colors into it's own node. This would result in explicit connections and message passing between articles.

Right now, the categorical features are encoded as an integer 0-n, where n is the number of unique categorical values. (eg. 0-male, 1-female).
That's not ideal for any NN, so we'll need to turn this into basically a "word embedding" with pytorch's Embedding layer.
We'll need to do this separately for all the different categorical features we have.
The Embedding layer in pytorch is a dictionary, where the parameters are learnable, so it'll need to be part of the model, not the pre-processing step.