Do large negative user biases mean anything?

I have created a LightFM model that recommends users movies and TV series based on their viewing history. The model works reasonably well, with a test AUC score around 0.75 and a test precision@K=15 score around 0.05. Subjectively speaking, the generated recommendations are fine although a bit more generic/less diverse than I'd ideally want.

I was looking at the user biases, and noticed that they're all large and negative (with a mean around -175) both for users with few and for users with many interactions. The item biases (and user/item embeddings) look "normal", lying in a range between roughly -0.5 and 0.5.

Does this mean anything, e.g. that the model hasn't converged properly?


Some additional details about the data and model: each item and user has several features indicating things like genre (item) and geographical location (user). The model parameters are as follows: no_components=37, max_sampled=30, loss='warp', learning_rate=0.05, item_alpha=1e-4, user_alpha=1e-4, epochs=430.

No source file or test is identified. Start by reviewing the LightFM model configuration and the reported user/item bias outputs, then determine whether the large negative user biases indicate a convergence or interpretation problem; done means documenting a supported explanation or a reproducible defect.

Written by the indexing model from the issue text.