B1483
Title: Uncertainty quantification in multi-omics data analysis and beyond
Authors: Florian Buettner - German Cancer Research Center (DKFZ) Heidelberg (Germany) [presenting]
Abstract: With model trustworthiness being crucial for sensitive real-world applications, practitioners are focusing more on improving the uncertainty awareness of machine learning models. This raises the need to quantify and improve the quality of predictive uncertainty, ideally via a dedicated metric. An uncertainty-aware model should give probabilistic predictions representing the true likelihood of events depending on the prediction. To quantify the extent to which this condition is violated, calibration errors have been introduced, and post-hoc recalibration methods are commonly used to improve them. However, estimators of calibration errors are usually biased and inconsistent. The framework of proper calibration errors is introduced, which gives important guarantees and relates every calibration error to a proper score. The improvement of an injective recalibration method w.r.t. a proper calibration error is reliably estimated via its related proper score. An orthogonal way of quantifying model-intrinsic uncertainties is via a Bayesian approach. Here, MuVI is presented, a novel multi-view latent variable model based on a modified horseshoe prior for modelling structured sparsity. This facilitates the incorporation of limited and noisy domain knowledge, thereby allowing for an analysis of multi-view data under uncertainty.
