Das Kolloquium findet im Freihausgebäude der TU Wien (Adresse: Wiedner Hauptstraße 8-10, 1040 Wien) statt.

Unsere Vorträge werden über E-Mail ausgeschrieben. Um diese Ankündigungen zu erhalten senden Sie eine E-Mail an sympa@list.tuwien.ac.at mit dem Betreff "sub swm-kolloquium" um in die E-Mail-Liste eingefügt zu werden.

(Abmeldungen von dieser Liste sind jederzeit möglich, indem Sie eine E-Mail an sympa@list.tuwien.ac.at mit dem Betreff "unsub swm-kolloquium" senden.)

Upcoming Seminars

October 29, 2026, 4.30 pm, Freihaus Seminar Room DB yellow 05A (yellow area, 5th floor) Wiedner Hauptstraße 8-10, 1040 Wien

Dr. Ruth Pfeiffer, öffnet eine externe URL in einem neuen Fenster, National Cancer Institute, National Institutes of Health (USA)
Title: Sufficient Dimension Reduction for Compositional Data

Abstract: Compositional data arise in many scientific fields and consist of vectors whose components satisfy a constant-sum constraint. A prominent example is microbiome data generated by high-throughput technologies to characterize microbial communities in biological samples, such as saliva or stool. We develop likelihood-based sufficient dimension reduction (SDR) methods to identify linear combinations of compositional predictors (X) that retain all information relevant to modeling and predicting an outcome (Y). Our approach uses the inverse distribution of (X) given (Y), considering several pairwise graphical models (pGMs), including normal and Poisson models, to capture complex dependencies among components.Microbiome data often contain more zeros than standard exponential-family pGMs can accommodate. To address this limitation, we introduce a class of zero-inflated pGMs with exponential-family distributions that accommodate excess zeros. We derive the corresponding first-order sufficient reductions and estimate them using penalized pseudo-likelihood, with a hierarchical penalty that incorporates the graphical model structure and enables variable selection. The resulting estimators of the reduction are consistent, and the approach accommodates continuous and categorical outcomes. We evaluate the methods in simulations involving normal, Poisson, and truncated Poisson graphical models with excess zeros and apply them to microbiome data from the American Gut Project. The proposed models provide robust variable selection, and the zero-inflated Poisson pGM achieves predictive performance comparable to or better than that of existing methods for microbiome data analysis.

Past Seminars