Reproduce “Learning Conditional Variational Autoencoders with Missing Covariates” on OpenResearch

About this paper

Siddharth Ramchandran, Gleb Tikhonov, Otto Lönnroth, Pekka Tiikkainen, Harri Lähdesmäki

A probabilistic method is introduced for learning conditional Variational Autoencoders (CVAEs) and Gaussian Process (GP) prior VAEs from datasets with missing auxiliary covariates. The method improves prediction accuracy for dependent variables and imputation accuracy for missing covariates, outperforming heuristic imputation baselines on synthetic and real-world clinical trial datasets.

Key insights

• Explicitly modeling and marginalizing the uncertainty of missing auxiliary covariates within the VAE's probabilistic framework leads to more robust and accurate learning than relying on simple imputation techniques.
• Decomposing the ELBO allows for a principled separation of the KL divergence terms related to the primary latent variables and the missing covariates, enabling integrated inference.
• The approach can be effectively scaled to advanced GP prior VAEs, extending their utility to complex temporal and longitudinal data where both correlations and missing covariates are prevalent.

Summary by alphaXiv.