Abstract
This paper provides a comprehensive analysis of variational inference inlatent variable models for survival analysis, emphasizing the distinctivechallenges associated with applying variational methods to survival data. Weidentify a critical weakness in the existing methodology, demonstrating how apoorly designed variational distribution may hinder the objective of survivalanalysis tasks--modeling time-to-event distributions. We prove that the optimalvariational distribution, which perfectly bounds the log-likelihood, may dependon the censoring mechanism. To address this issue, we propose censor-dependentvariational inference (CDVI), tailored for latent variable models in survivalanalysis. More practically, we introduce CD-CVAE, a V-structure VariationalAutoencoder (VAE) designed for the scalable implementation of CDVI. Furtherdiscussion extends some existing theories and training techniques to survivalanalysis. Extensive experiments validate our analysis and demonstratesignificant improvements in the estimation of individual survivaldistributions.