Abstract: Longitudinal datasets naturally arise in a wide variety of fields and applications, including biomedicine, healthcare, psychology, consumer behaviour, and many others. In this talk, I will present our recent work on developing deep and non-parametric methods for longitudinal data analysis. Our methods build on interpretable additive Gaussian processes (GP) and extend them to e.g. account for temporal uncertainty in covariates, heterogeneity of covariate effects across subjects, appropriate observation models, and propose an accurate covariate relevance assessment method. For high-dimensional data, our method can be formulated as a longitudinal variational autoencoder (L-VAE) with a multi-output additive GP prior. We derive a new and tighter KL divergence upper bound for such GPs, and devise a mini-batch compatible learning method for L-VAE that exploits the natural gradients. L-VAE can simultaneously accommodate both time-varying shared and random effects, produce structured low-dimensional representations, disentangle effects of individual covariates or their interactions, and achieve highly accurate predictive performance. I will demonstrate these methods on various biomedical as well as synthetic data sets.

These are joint works with Juho Timonen, Siddharth Ramchandran, Gleb Tikhonov, Aki Vehtari and others, and will be published in Bioinformatics and AISTATS in 2021.
https://academic.oup.com/bioinformatics/advance-article/doi/10.1093/bioinformatics/btab021/6104850
https://arxiv.org/pdf/2006.09763.pdf

Speakers:  Harri Lähdesmäki

Affiliation: Aalto University

Place of Seminar:  Zoom (Available afterwards on Youtube)
Meeting ID: 677 9168 1514
Passcode: 967449