Marformer: A Transformer for Predicting Missing Data Distributions
Marformer predicts conditional marginals for missing variables directly, without building the full joint distribution.
The paper frames those marginals as the quantity needed for Bayes risk and value-of-information decisions under incomplete data. Its Transformer takes any observed subset and produces the missing-variable distributions in one forward pass. The authors test it on synthetic Bayesian networks, discretized Gaussians, structured annotation data, and one real annotation dataset. They report that it matches or beats the evaluated baselines in those settings and runs substantially faster than the generative ones. ArXiv · AI/CL/LG's note
The paper frames those marginals as the quantity needed for Bayes risk and value-of-information decisions under incomplete data. Its Transformer takes any observed subset and produces the missing-variable distributions in one forward pass. The authors test it on synthetic Bayesian networks, discretized Gaussians, structured annotation data, and one real annotation dataset. They report that it matches or beats the evaluated baselines in those settings and runs substantially faster than the generative ones. ArXiv · AI/CL/LG's note
score 4