MatrixFormer: A Foundation Model for Matrix Completion
MatrixFormer predicts distributions for all missing matrix entries in one pass.
The paper presents a matrix-native transformer for matrix completion, instead of treating missing values as separate entry-level predictions. It is pretrained only on synthetic low-rank and latent-factor matrices with varied missingness patterns. With the same weights, the authors report zero-shot competitive results across panel-data causal inference, benchmark-score completion, tabular imputation, and recommender-system completion. ArXiv · AI/CL/LG's note
The paper presents a matrix-native transformer for matrix completion, instead of treating missing values as separate entry-level predictions. It is pretrained only on synthetic low-rank and latent-factor matrices with varied missingness patterns. With the same weights, the authors report zero-shot competitive results across panel-data causal inference, benchmark-score completion, tabular imputation, and recommender-system completion. ArXiv · AI/CL/LG's note
score 6