Robust learning under natural rating noise for collaborative filtering

In intelligent information systems that rely on explicit-feedback recommendation, observed ratings are noisy, subjective proxies for user preference. Most prior work on natural noise in collaborative filtering handles this problem through data cleansing. That strategy separates noise handling from model training and may discard informative observations. This paper instead studies robust learning directly in the training objective for explicit-rating prediction. We reformulate rating prediction as a multiclass problem under label noise. We adapt Forward Correction and Backward Correction to collaborative filtering and introduce an anchor-based heuristic for estimating the noise transition matrix when the corruption process is unknown. The approach is evaluated on MovieLens 100k, CiaoDVD, and Yahoo! Music under controlled uniform and pairwise corruption. It compares corrected losses with cleansing methods, robust-learning baselines, and classical collaborative-filtering predictors in a shared explicit-rating setting. Across datasets and corruption levels, Forward Correction with the proposed estimator delivers the most consistent balance between predictive quality and robustness. The estimated matrix also appears to capture part of the noise the datasets already carry, in addition to the injected corruption, although that evidence is necessarily indirect: no public explicit-feedback dataset annotates which ratings are inconsistent. These findings suggest that, for explicit-rating prediction under controlled corruption, robust learning is a strong alternative to pre-processing-centered noise handling in collaborative filtering.