Dr. Hsin-Hsiung 'Bill' Huang, professor in UCF's School of Data, Mathematical and Statistical Sciences, will speak on "Bayesian Sparse Mixed-Response Species Distribution Models with Spatial Dependence" at this week's seminar.

Abstract: Ecological monitoring programs increasingly record continuous, binary and count responses at common sites. We develop a Bayesian species distribution model for such mixed ecological responses, using a shared row-sparse coefficient matrix, Pólya–Gamma augmentation for binary and negative-binomial components, Gaussian updates for continuous components, and a low-rank representation of spatial or spatio-temporal structure. Raw environmental variables are expanded through a deterministic maximum-entropy-style feature dictionary before joint shrinkage is applied. The model preserves response-family-specific likelihoods while allowing common environmental structure and residual dependence across responses. Controlled simulations show that the one-step and screened two-step estimators recover sparse coefficient rows and improve binary discrimination relative to response-wise sparse regressions when the data-generating mechanism contains shared sparsity and cross-response dependence. Heat maps, support-recovery summaries and residual maps diagnose sensitivity to spatial rank, count calibration and screening choices. The ecological analyses clarify the intended use. FishGlob motivates the full mixed-response survey problem, family-specific diagnostics and dependence visualization. The African elephant occurrence-background benchmark shows why the Bernoulli special case must use the full spatial-sparse predictor rather than an environment-only sparse ablation: a strong spatial gradient cannot be represented adequately by sparse environmental features alone. After adding the spatial-basis component and training-fold probability calibration, the proposed Bernoulli implementation has the strongest blocked cross-validation area under the receiver operating characteristic curve, Brier score and log score among the fitted elephant comparators. The method is most useful for high-dimensional mixed-response learning with interpretable shared sparsity and explicit spatial adjustment.

Speaker Bio: Dr. Hsin-Hsiung “Bill” Huang is a professor in statistics and data science at the University of Central Florida. He earned his doctoral degree in statistics from the University of Illinois at Chicago, a master's in statistics from the Georgia Institute of Technology, a master's in mathematics and dual B.A./B.S. degrees in economics and mathematics from National Taiwan University. Dr. Huang’s research interests include Bayesian ultrahigh-dimensional variable selection, matrix- and tensor-variate regression, clustering, classification, and dimension reduction, with a focus on developing new statistical methods for complex, large-scale data. His recent projects encompass spatiotemporal modeling, novel neuroimaging reconstruction algorithms and threat detection. His work has been supported by multiple grants from the National Science Foundation’s Algorithms for Threat Detection program, where he serves as principal investigator (including NSF DMS-1924792 and DMS-2318925), as well as by the National Institutes of Health, where he is a co-investigator on a National Institute of Neurological Disorders and Stroke (NINDS) R01 award (1R01NS133094-01A1).