Determine whether a third latent disagreement dimension is genuine

Determine whether the marginal improvement from adding a third latent factor to the Community Notes bridging matrix-factorisation model reflects a genuine third dimension of rater disagreement or residual overfitting, using a full-scale rank-3 fit under the same held-out validation protocol.

Background

The study compares rank-0, rank-1, rank-2, and rank-3 versions of the Community Notes latent-factor model. The second dimension produces a substantial improvement in held-out rating prediction, whereas the third produces only a marginal gain of 0.0008 RMSE, or 0.3% relative improvement. Variance decomposition on the newer snapshot gives the three dimensions similar shares, but predictive evidence for the third dimension is weak.

The authors therefore retain a two-dimensional model as the primary specification, while noting that a complete rank-3 fit on the same full-scale data and under the same held-out protocol would be needed to distinguish a real third axis of disagreement from overfitting. The unresolved issue concerns the dimensionality of the latent viewpoint space that the production bridging mechanism should represent.