Notes
🔍

Rating Prediction

The first way anyone framed recommendation was as filling in a blank: here is a grid of users and the star ratings they gave, nearly all of it empty — predict the missing numbers. That framing is rating prediction, and it shaped a decade of the field.

What the task is
Given a user-item pair the user has not rated, output the score they would give. It is regression over a sparse matrix, trained on observed cells and scored on held-out ones with RMSE or MAE. The classical solutions belong to this framing: a global mean plus user and item biases as the baseline, then Funk SVD and its descendants learning latent factors whose dot product reconstructs the rating. The Netflix Prize was a rating-prediction contest, and MovieLens is still its reference dataset.

Why the field moved on

  • Real systems collect almost no ratings. Clicks, plays and purchases are plentiful; stars are rare and self-selected.
  • Low error on rated items does not imply a good list — a model can predict every rating within half a star and still put nothing worth watching in the top ten.
  • Only items the user chose to rate are ever scored, so the rest of the catalogue is invisible to the metric.

Where it still fits
Where explicit scores genuinely exist and the number itself is shown to a person: review platforms, internal quality scoring, editorial tools. Everywhere else the output is consumed as an ordered list, so top-N recommendation is the framing that matches the product, and a squared error on an item nobody will ever reach is a cost the model should not be paying.

Rating prediction frames recommendation as regression: estimate the explicit score a user would give an unseen item, filling empty cells of the user-item matrix and scoring the result with RMSE or MAE. It defined the classical era of biases plus matrix factorization, and still fits products that display a predicted score. Where feedback is implicit, accurate ratings do not guarantee a good list, so ranking formulations replace it.

Picture a huge spreadsheet: every customer down the side, every film across the top, and a star rating in a cell whenever someone bothered to leave one. Almost every cell is empty.

Rating prediction is the job of guessing what those empty cells would say. Work out that you would give a particular film four stars, and the system knows whether to suggest it. This is how recommendation was first thought about, back when websites asked you to rate things you had finished. Most services now watch what you click instead, so the question has quietly shifted from what score you would give to what you would pick.