BEGIN:VCALENDAR
X-WR-CALNAME:Webnotice (Statistics & Actuarial Science)
X-WR-CALDESC:Webnotice (Statistics & Actuarial Science) at University of Waterloo
X-PUBLISHED-TTL:PT60M
PRODID:-//UW-Webnotice/NONSGML 0.1//EN
VERSION:2.0
BEGIN:VEVENT
DTSTAMP:20210916T200000Z
UID:2021_ad4e3b2c2f7960404056ff967b0f4a3d.wnotice@math.uwaterloo.ca
DTSTART:20210916T200000Z
DTEND:20210916T210000Z
SUMMARY:Spatial meshing for general Bayesian multivariate models (Statistics and Biostatistics Seminar)
LOCATION:Virtually on Zoom
DESCRIPTION:Spatial meshing for general Bayesian multivariate models\nMichele Peruzzi, Duke University\n\nQuantifying spatial associations in multivariate geolocated data of different types is achievable via spatial random effects in a Bayesian hierarchical model, but severe computational bottlenecks arise when spatial dependence is encoded as a latent Gaussian process (GP) in the increasingly common big data settings on which we focus. The scenario worsens in non-Gaussian models because the reduced analytical tractability leads to additional hurdles to computational efficiency. We introduce a methodology for efficiently computing multivariate Bayesian models of spatially referenced data in which the likelihood or the latent process (or both) are not Gaussian. We exploit the advantages of spatial processes built via directed acyclic graphs, in which case the spatial nodes enter the Bayesian hierarchy and lead to an intuitive algorithm for posterior sampling via routine Markov chain Monte Carlo (MCMC) methods. Our methods compare favorably to state-of-the-art MCMC-free methods, and enable scalable posterior sampling of all unknowns even with data in the millions, as demonstrated on a bivariate spacetime application with misaligned satellite imaging with n > 2.5 * 10^6. All methods are available as part of R package 'meshed', available on CRAN. \n\nHosted on Zoom: https://uwaterloo.zoom.us/j/98471150483
END:VEVENT
BEGIN:VEVENT
DTSTAMP:20210923T200000Z
UID:2021_0c74d70575c3e01f57130f8742613116.wnotice@math.uwaterloo.ca
DTSTART:20210923T200000Z
DTEND:20210923T210000Z
SUMMARY:Our Recent Development on Cost Constraint Machine Learning Models (Statistics and Biostatistics Seminar)
LOCATION:Virtually on Zoom
DESCRIPTION:Our Recent Development on Cost Constraint Machine Learning Models\nHaoda Fu, Eli Lilly and Company\n\nSuppose we can only pay $100 to diagnose a disease subtype for selecting best treatments. We can either measure 10 cheap biomarkers or 2 expensive ones. How can we pick the optimal combinations to achieve highest diagnostic accuracy? \n\nThis is a nontrivial problem. For a special case, as each variable costs the same, the total cost constraint will be reduced to an L0 penalty which is the best subset selection problem. Until recently, there is no good solution even for this special case. Traditional algorithms can only solve up to ~35 variables for best subset selections. Thanks to the algorithms breakthrough in the field of optimization research. We have modified and extended a recently developed algorithm to handle our cost constraint problems with thousands of variables. \n\nIn this talk, we will talk about the background of this problem, methods development, theoretical results. We will also show you an impressive example on dynamic programming. It will tell a story on how algorithms can make a difference on computing. I hope that through this talk, you can feel the modern statistics which combined computer science, statistics, and algorithms.
END:VEVENT
BEGIN:VEVENT
DTSTAMP:20210924T140000Z
UID:2021_01fe8f208531e525ec6ef3ea7f706982.wnotice@math.uwaterloo.ca
DTSTART:20210924T140000Z
DTEND:20210924T150000Z
SUMMARY:Conditional mean risk sharing in the individual model for dependent losses (Actuarial Science and Financial Mathematics Seminar)
LOCATION:Virtually on Zoom
DESCRIPTION:Conditional mean risk sharing in the individual model for dependent losses\nChristian Y. Robert, École nationale de la statistique et de l'administration économique Paris (ENSAE)\n\nThe conditional mean risk sharing rule for insurance losses, as defined by Denuit and Dhaene (2012), provides actuaries with an effective method to deal with collaborative insurance pools or mutual aid funds. This talk considers the case when participants losses are dependent. Some examples of multivariate distributions for which it is possible to obtain explicit formulas for the individual contributions are first presented. Two frameworks are then discussed where analytical formulas are derived and numerical or simulation approaches are developed to evaluate the contributions: the conditionally independent risks model and the model with graphical dependencies. \n\nHosted on Zoom: https://uwaterloo.zoom.us/j/97141496491
END:VEVENT
END:VCALENDAR