Incorporating item and user biases in LightFM embeddings for accurate recommendations in Vespa

Nobody has claimed this yet.

I am working on deploying a recommender system using LightFM and Vespa. I have trained a LightFM model and saved the item embeddings in Vespa for deployment. However, I am facing an issue with prediction accuracy as I can only use the dot product of user and item embeddings for prediction in Vespa. The LightFM model's prediction formula includes item_biases and user_biases, which are crucial for accurate recommendations (user_e . item_e + user_b + item_b). Unfortunately, Vespa does not support entering bias information for user and item embeddings, resulting in reduced prediction accuracy.

I would appreciate any guidance or suggestions on how to train the model without biases or how to collect the information of embeddings and biases together, so that I can make accurate predictions in Vespa.
Unfortunately, ignoring biases or concatenating them to embeddings led to worsened prediction accuracy.
Thank you for your help.

No source file, test, or entry point is named. Start by reviewing LightFM's prediction formula and the Vespa embedding limitation described in the report; done would require an agreed, accurate way to represent or apply user and item biases for Vespa predictions.

Written by the indexing model from the issue text.