RecBole
RecBole copied to clipboard
[🐛BUG] Context-aware recommenders not properly embedding float sequences.
Hello! Thank you for this library.
Describe the bug I am running several context-aware models using several float-sequence item features. However, the results that I get are for each model always the same, irrespective of the feature loaded.
While inspecting the code of Context Aware recommenders, I noticed that float sequences are always embedded with a number of embeddings of float_seq_field_dim
. However, float_seq_field_dim
is always 1 for float sequences, as can be seen here and here. This means that every item is assigned the same embedding. This would explain why I get the same result, no matter which item feature I am loading.
Expected behavior The model performance should vary depending on the float seq feature used.
My guess on the origin of the problem float sequences are always embedded with a number of embeddings of 1, which means that every item is assigned the same embedding. This would explain why I get the same result, no matter which item feature I am loading.
Are my hypotheses correct? If so, I think I found a bug :)
Best, Marta