Longitudinal Latent Diffusion Models
Keywords: generative AI, high-dimensional data, longitudinal data, diffusion models, variational autoencoders, latent representations
TL;DR: A new generative method for high-dimensional longitudinal data, based on latent diffusion models and a geometric perspective on variational autoencoder latent space.
Abstract: Longitudinal data are crucial in several fields, but collecting them is a challenging process, often hindered by concerns such as individual privacy. Extrapolating in time initial trajectories or generating fully synthetic sequences could address these issues and prove valuable in clinical trials, drug design, and even public policy evaluation. We propose a generative statistical model for longitudinal data that links the temporal dependence of a sequence to a latent diffusion model and leverages the geometry of the autoencoder latent space. This versatile method can be used for several tasks - prediction, generation, oversampling - effectively handling high-dimensional data such as images and irregularly-measured sequences, needing only relatively few training samples. Thanks to its ability to generate sequences with controlled variability, it outperforms previously proposed methods on datasets of varying complexity, while remaining interpretable.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2595
Loading