WIMS2011 Costa
Paper by Alberto Costa and Fabio Roda (LIX, École Polytechnique) proposing a method that reformulates collaborative recommender-system prediction as an information-retrieval task: users = documents, items (movies) = terms, active user = query, and ratings used as term weights. They test this approach on the GroupLens 100K dataset using two IR models (LSPR and a vector-space model), evaluate predictions with RMSE under leave-n-out splits, and report modest improvements over community-average baseline (LSPR ~8.4% improvement, vector space ~7.6%). The work is positioned as an exploratory empirical investigation and outlines plans for further tuning and comparisons to KNN, Slope-One, BM25, Terrier, and ensemble strategies.
Questions this document answers
GroupLens 100K (943 users, 1682 movies, ~100,000 ratings, ratings 1–5).
Least Spectral Power Ranking (LSPR) and a basic vector space model.
Root Mean Squared Error (RMSE) computed on leave-n-out (80/20) splits; predictions rounded to nearest integer.
Approximately 8.4% improvement (reported average).