Isn't this basically a de-biasing problem? Treat each rider’s ratings as a random variable with its own mean μᵤ and variance σᵤ², then normalize. Basically compute z = (r – μᵤ)/σᵤ, then remap z back onto a 1–5 scale so “normal” always centers around ~3. You could also add a time decay to weight recent rides higher to adapt when someone’s rating habits drift.

Has anyone seen a live system (Uber, Goodreads, etc.) implement per-user z-score normalization?