Programación del seminario: Año 2022

•  Viernes, 10 de junio de 2022. Hora: 12:00

Lorenzo Cappello. Department of Economics, Universitat Pompeu Fabra

An efficient coalescent model for heterochronously sampled data

Inicio

An efficient coalescent model for heterochronously sampled data

INVITADO: Lorenzo Cappello

IDIOMA: inglés

LUGAR:  Seminari EIO, ETSEIB (ed. Eng. Industrial), Planta 6, Campus Sud. UPC, Av. Diagonal 647, Barcelona

FECHA: Viernes, 10 de junio de 2022. Hora: 12:00

RESUMEN: The observed sequence variation at a locus informs about the evolutionary history of the sample and past population size dynamics. The Kingman coalescent is used in a generative model of molecular sequence variation to infer evolutionary parameters. However, it is well understood that inference under this model does not scale well with sample size. In the talk, we will discuss recent works based on a lower resolution coalescent process, the Tajima coalescent, to model longitudinal samples. While the Kingman coalescent models the ancestry of labeled individuals, we will model the ancestry of individuals labeled by their sampling time. Some interesting statistical questions arise when trying to fit this model, and we will discuss them. The need to model longitudinal samples was motivated by applications (e.g., ancient DNA and RNA from rapidly evolving pathogens like viruses). Part of the talk will discuss some of these and possible future directions.

ACERCA DEL AUTOR:  Lorenzo is an assistant professor in the Statistics group at Universitat Pompeu Fabra, Department of Economics, and affiliated professor at the Barcelona School of Economics. Previously, he was a postdoctoral scholar in the Department of Statistics at Stanford University. He earned his Ph.D in Statistics in 2018 from the Department of Decision Sciences at the Bocconi University. His research focuses on providing computationally manageable methods and scalable algorithms that adapt to big data problems. His research is motivated by specific questions that arise in applications - in particular population genetics and applied mathematics - and address them using statistical methods. For more information on his research interest, see https://lorenzocapp.github.io/

Inicio

Una científica en temps de pandèmia: les matemàtiques com a eina clau en el control de la COVID-19

INVITADA: Clara Prats

IDIOMA: Catalán

LUGAR:  Sala de Actos Manuel Martí Recober (FIB), Planta 0, Edificoi B6, Campus Norte. UPC, Barcelona.

FECHA: Martes, 14 de junio de 2022. Hora: 12:30

RESUMEN: Hace dos años, la pandemia de COVID-19 puso nuestras vidas patas arriba. Desde el mundo de la ciencia, a menudo demasiado cerrado, nos encontramos con la necesidad de dar un paso adelante (o muchos!) y ponernos al servicio de la sociedad en plena emergenica. Durante este tiempo, las matemáticas han sido una herramienta imprescindible para analizar los datos epidemiológicos, y transformar una situación compleja en un conjunto de indicadores objetivos, válidos para comprender esta situación y dar soporte a la toma de decisiones. También ha sido una herramienta muy importante para entender la evolución de la pandemia, con sus sucesivas olas y el efecto de las distintas medidas en su modulación. Finalmente, los modelos matemáticos han sido imprescindibles en el análisis de dinámicas de propagación y en las predicciones a corto y medio plazo. Las matemáticas, pues, han sido una pieza esencial en los equipos multidisciplinares que han trabajado para la pandemia desde la ciencia y desde la gestión.

SOBRE LA AUTORA:  Clara Prats es licenciada en Física per la Universidad de Barcelona y és doctora en Física aplicada y simulación en ciencia por la UPC.  Actualmente es profesora titular en el departamento  de Física,  investigadora del grupo  Computational Biology and Complex Systems (BIOCOM-SC) y responsable de modelos computacionales en el Centre for Comparative Medicine and Bioimaging (CMCiB) del Instituto Germans Trias i Pujol (IGTP). Su investigación se centra en el uso de modelos computacionales para el estudio de enfermedades infecciosas; principalmente, tuberculosis y COVID-19. Siguiendo  este enlace se puede acceder a información actualizada sobre la carrera de Clara.

Inicio

Model selection and model uncertainty quantification in survival models

INVITADA: María Eugenia Castellanos

IDIOMA: spanish

LUGAR:  Seminari EIO, ETSEIB (ed. Eng. Industrial), Planta 6, Campus Sud. UPC, Av. Diagonal 647, Barcelona

FECHA: Miércoles, 29 de junio de 2022. Hora: 12:00

RESUMEN:We consider covariate selection and the ensuing model uncertainty aspects in the context of survival regression, specifically for log-normal and Cox regression. The perspective we take is probabilistically handled within a Bayesian framework. One of the critical elements in variable/model selection is choosing a suitable prior for model parameters. Here we derive the so-called conventional prior approach and propose a comprehensive implementation that results in an automatic procedure. Our simulation studies and real applications show improvements over existing literature. Finally, with our proposal is possible to make predictions about the survival function for a specific type of patient, incorporating all the uncertainty probabilistically. Trabajo conjunto con Gonzalo García-Donato y Stefano Cabras.

SOBRE LA AUTORA: María Eugenia Castellanos es catedrática de Estadística e Investigación Operativa en la Universidad Rey Juan Carlos de Madrid dónde imparte clases en diversos grados y masters, además de ocuparse actualmente de la Subdirección de Relaciones Internacionales e Investigación de la Escuela Técnica Superior de Ingeniería Informática. Anteriormente ha trabajado en la Universidad Miguel Hernández y la Universidad de Valencia. Su investigación se centra en la estadística bayesiana, en ámbitos como el contraste y selección de modelos, el análisis de supervivencia y modelos para valores extremos entre otros. Más información en: https://mecastellanos.wordpress.com/.

Inicio

Understanding complex predictive models through Ghost Variables

NVITADO: Pedro Delicado

IDIOMA: español

LUGAR:  Seminari EIO, ETSEIB (ed. Eng. Industrial), Planta 6, Campus Sud. UPC, Av. Diagonal 647, Barcelona

FECHA: Viernes, 11 de noviembre de 2022. Hora: 12:30

RESUMEN: Machine Learning (ML) models are more and more accurate in their predictions, many times at the cost of an increasing complexity, which is why we often refer to them as “black boxes”. A whole literature has recently appeared (Interpretable ML) whose purpose is to provide transparency and interpretability to predictive algorithms. In this context, we propose a procedure for assigning a relevance measure to each explanatory variable in a complex predictive model. Assuming that a test set is available, the individual relevance of a variable is computed by comparing the model predictions for the test set with those given for a modified test set, in which the variable of interest is substituted by its “ghost variable”, defined as the prediction of this variable as a function of the rest of explanatory variables. We illustrate our proposal with simulated examples and the analysis of a real data set on rental housing. (Based on a joint paper with Daniel Peña: TEST, 2022, DOI 10.1007/s11749-022-00826-x.)

ACERCA DEL AUTOR:  Pedro Delicado is Professor of Statistics at UPC. His research activity has been mainly dedicated to Functional Data Analysis (focusing on dimensionality reduction and spatial dependency), but in recent years he is also interested in exploring the links between Statistics and Machine Learning, with special interest in the interpretability of predictive models. For more information on his research interest, see  http://www-eio.upc.edu/~delicado/index.html

Inicio