Training on SCARF Embeddings results in worse accuracy

[golnaz-amii]

Hi!

I'm new to SSL for tabular data, and was wondering why in example/example.ipynb model performance is worse on the embedding than raw features, and if there are ways to debug/finetune/tune the hypers so that SCARF actually helps

Thank you