Data Analysis and Statistical Science

Quetelet seminars 2016

Quetelet seminars 2016

 

Probabilistic Index Models

Prof. Olivier Thas
Ghent University

Wednesday , December 21, 2016 , 13h00
E1.015 , Campus Coupure, building E, first level

Abstract:

Modelling uncertainty in ordinal data analysis

Dr. Rosaria Simone
University of Naples Federico II, Department of Political Sciences

Tuesday , October 25, 2016 , 13h00
Practicumzaal A , Campus Dunant, ground floor, Dunantlaan 1, 9000 Gent

Abstract:

CUB mixture models refer to a statistical framework for rating data allowing for a direct measurement of uncertainty. In survey analysis, the process of translation of the latent perception about the investigated item into an ordinal evaluation produces unobserved fuzziness that blur the expressed preferences. According to CUB model paradigm, the data generating process is untangled and explained as a blend of real beliefs and indecision: a mixture of a feeling component and an uncertainty component. Also respondents specific variables can be included in the model to enhance the different contributions of subjects’ covariates on response patterns. The benchmark CUB model can be suitably adapted to encompass typical phenomena arising in ordinal data analysis, such as the presence of inflated category and overdispersion. The talk will present this perspective by integrating methodology with empirical evidence. Future research developments and open issues will be outlined.

A gentle introduction to Bayesian evaluation of informative hypotheses

Dr. Herbert Hoijtink
Universiteit Utrecht

Tuesday , October 18, 2016 , 13h00
Practicumzaal A , Campus Dunant, ground floor, Dunantlaan 1, 9000 Gent

Abstract:

This presentation will introduce Bayesian evaluation of informative hypotheses. Using an empirical example from the area of stress research, it will be shown that classical hypotheses testing by means of the p-value is often unable to provide answers to the research questions of interest. Subsequently, it will be elaborated that informative hypotheses can be used to represent the ideas that researchers have about their area of interest. Finally, it will be shown how the Bayes factor can be used to evaluate these hypotheses.

Measuring Sample Quality with Stein's Method

Dr. Lester Mackey
Stanford University

Tuesday , September 27, 2016 , 13h00
Room V3 , S9, Campus Sterre, 9000 Gent

Abstract:

To improve the efficiency of Monte Carlo estimation, practitioners are turning to biased Markov chain Monte Carlo procedures that trade off asymptotic exactness for computational speed. The reasoning is sound: a reduction in variance due to more rapid sampling can outweigh the bias introduced. However, the inexactness creates new challenges for sampler and parameter selection, since standard measures of sample quality like effective sample size do not account for asymptotic bias. To address these challenges, we introduce a new computable quality measure based on Stein's method that quantifies the maximum discrepancy between sample and target expectations over a large class of test functions. We use our tool to compare exact, biased, and deterministic sample sequences and illustrate applications to hyperparameter selection, convergence rate assessment, and quantifying bias-variance tradeoffs in posterior inference.

Workshop Flexible Statistical Modeling

 

Thursday-Friday , September 15-16, 2016
, Het Pand, Ghent University, Onderbergen 1, 9000 Gent

Abstract:

Stimulate research in flexible modeling. Bring together researchers in the field to share their latest results. Identify interesting new problems and applications. Foster dialogue and future research collaborations.

Next generation image processing

Dr. Lukasz Kidzinski
Ecole polytechnique fédérale de Lausanne

Thursday , May 26, 2016 , 11h30
Room V3 , Building S9, Campus Sterre, Krijgslaan 281, 9000 Gent

Abstract:

In recent decades, the most popular approach to image processing has involved treating an image as a two-dimensional signal and applying signal processing filters. A large number of tools can be created this way and therefore it became the basis for modern image processing software. However, the scope of these techniques is limited by the fact that filters need to be designed manually. The recent advancements in image recognition and convolutional neural networks offer an alternative. Given large amount of images, filters for information extraction can be learned from the data. The algorithm recently introduced by Gatys et al. shows that these filters can also be successfully used for image generation. This discovery opens doors to the next generation image processing, exemplified by DeepArt.io software.

Introduction to nlsem: An R Package for Estimating Nonlinear, Structural Equation Mixture Models

Nora Umbach
University of Tuebingen

Friday , April 29, 2016 , 13h45
practicumzaal A , Campus Dunant, Henri Dunantlaan 1, 9000 Gent

Abstract:

An R package for fitting nonlinear structural equation models using normal mixture distributions has recently been published on CRAN. Implemented in this package are four different model approaches: The LMS and the QML approach, which allow for two-way interactions and quadratic effects in structural models with normally distributed predictor variables. The Structural Equation finite Mixture Model (STEMM or SEMM) approach (Bauer, 2005; Jedidi, Jagpal, & DeSarbo, 1997), which uses mixtures of linear structural equation models to deal with nonnormality. The recently proposed Nonlinear Structural Equation Mixture Model approach (NSEMM; Kelava, Nagengast, & Brandt, 2014). Here, interactions and quadratic terms as well as latent mixtures can be modeled, which allows for a separation of nonnormality of the latent predictors and nonlinearity of latent relationships. The user interface of the nlsem package and its main functions will be presented, along with some empirical examples. Limitations and future features of the nlsem package will be discussed.

Regression models for indirectly measured responses

Jan De Neve
Ghent University

Friday , April 29, 2016 , 15h00 - 16h00
Auditorium A , Campus Dunant, Henri Dunantlaan 1, 9000 Gent

Abstract:

The Center for Statistics and the Department of Data Analysis have the pleasure to invite you to the introductory lecture of our colleague Jan De Neve who started this academic year as a UGent Tenure Track professor in the discipline of Data-analysis for the behavioural sciences.

The lecture will be accessible to a wide audience and is followed by a reception.

Reception: start at 16h.
Venue: Practicumzaal A, Henri Dunantlaan 1, Campus FPPW.

PLEASE CONFIRM YOUR ATTENDANCE AT THE RECEPTION BEFORE APRIL 27 VIA THE LINK BELOW.

Inaugural lecture of Christophe Ley

Christophe Ley
Ghent University

Friday , March 25, 2016 , 11h30
Auditorium A2 , building S9, Campus De Sterre, Krijgslaan 281, Gent

Abstract:

The department of Applied Mathematics, Computer Science and Statistics is delighted to invite you to the "inaugural lecture" of our new colleague Christophe Ley, who started this year at Ghent University as a tenure track professor in Mathematical Statistics.

Title of his lecture: "Various aspects of mathematical statistics".

A new honorary doctor in statistics

 

Thursday, Friday , March 17-18, 2016

Abstract:

The Ghent University Faculty of science and Center for Statistics proudly present a new honorary doctor in statistics, 2 visionary speakers and a public debate.

Thursday, March 17, 17:00-19:30

Statistics and big data: Friend or foe?

Place: Campus Ledeganck, Auditorium 5, K.L. Ledeganckstraat 35, Gent.

17:00: Louise Ryan, University of Technology, Sydney (UGent honorary doctor 2016): "Statistics and Big data"

18:00 Drinks

18:30 Peter Diggle, president of the Royal statistical Society: "Statistics: a Data Science for the 21st Century"

Friday, March 18, 10:30-12:00

Facing the facts for a better world

Place: Faculty of Economics and Business Administration, Auditorium Quetelet, Tweekerkenstraat 2, 9000 Gent (Site Sint-Pietersplein)

In a world eager for information, the public and hence the media cares more for anecdotes than for carefully gathered evidence it seems. One episode of violence towards women in Koln does more to change awareness than 50.000 cases of domestic violence reported in Belgium per year. Should it? How are the numbers in the wake of the refugee crisis presented and what are the consequences? Today, carefully gathered statistics fall additionally into the shadows of sexy `big data analytics’. Whilst leading companies invest in data and technology to improve the health of their business, we wonder whether the health of society and indeed our personal health is best served by slow and careful science or fast and broad but personalized big data. What can drive meaningful changes for a healthy, prosperous and caring society? Where should academia and governance invest?

Join the debate with two honorary doctors 2016 and experts:

  • Prof. Michael Marmot, epidemiologist
  • Prof. Louise Ryan, biostatistician
  • A corona of experts including dr. Tijl De Bie, big data (UGent), prof. Ignaas Devisch, filosopher (UGent); prof. Peter Diggle, president of the Royal Statistical Society (UK); prof. Dirk De Bacquer, epidemiologist (UGent), dr. PauL Quataert, environmental biostatistician (INBO); prof. Joost Weyler, epidemiologist (UA)

Human-guided data analysis

Dr. Kai Puolamäki
Finnish Institute of Occupational Health

Tuesday , January 12, 2016 , 11h30 - ?
Room Shannon , Technicum B2, Sint-Pietersnieuwstraat 41, 9000 Gent

Abstract:

I will present our new project "Human-guided data analysis", its background, the scientific rationale, and the objectives. The project is funded by the Academy of Finland from September 2015 to August 2019.

We argue that in the field of explorative data analysis the human component has been neglected. In order to design data analysis methods suitable for human use then only using technical performance criteria, such as classification accuracy etc., is not enough. Our objective is to design data analysis methods that take the human cognition optimally into account in a principled way, allowing the computational methods to become an efficient extension of user's cognition in data exploration.

The main computational methodology used in the project is based on randomization and statistical significance testing approach. In randomization approach we can elegantly separate the design of algorithms and statistics. We can often use state-of-the-art machine learning methods, such as classifiers, without modification. Controlled random perturbations that we perform for the data allow us to, on the one hand, provide statistical guarantees, and on the other hand, allow user to feed in information for the computational process in the forms of constraints to randomization.

The abstract space provided by constraints to randomization can be viewed as "information space network" in which the user traverses during the explorative data analysis process. A challenge, tackled in the project, is to design the information space and its presentation so that it is efficiently traversable by the human user.

The project will result in deeper understanding of explorative data analysis process and computational methods to best support it. The Finnish Institute of Occupational Health (FIOH), which is the site of research, has many interesting scientific and medical data sets and use cases which we can use as examples during this project.

The project team at FIOH consists of Dr Kai Puolamäki (PI), Mr Andreas Henelius, Dr (psychology) Virpi Kalakoski, Dr (computer science) Emilia Oikarinen, and Dr (computer science) Antti Ukkonen. The project is related to our concurrent Revolution of Knowledge Work project, a large strategic research opening funded by Tekes - the Finnish Funding Agency for Innovation, see http://www.reknow.fi/

Selected references related to the topics of the talk:
Andreas Henelius, Kai Puolamäki, Henrik Boström, Lars Asker, Panagiotis Papapetrou. A peek into the black box: exploring classifiers by randomization. Data Mining and Knowledge Discovery, 28(5-6): 1503-1529, 2014. http://dx.doi.org/10.1007/s10618-014-0368-8

Jefrey Lijffijt, Panagiotis Papapetrou, and Kai Puolamäki. A statistical significance testing approach to mining the most informative set of patterns. Data Mining and Knowledge Discovery, 28(1):238–263, 2014. http://dx.doi.org/10.1007/s10618-012-0298-2

Kai Puolamäki, Panagiotis Papapetrou, and Jefrey Lijffijt. Visually controllable data mining methods. In IEEE International Conference on Data Mining Workshops 2010, 2010. http://dx.doi.org/10.1109/ICDMW.2010.141

Sami Hanhijärvi, Markus Ojala, Niko Vuokko, Kai Puolamäki, Nikolaj Tatti, and Heikki Mannila. Tell me something I don’t know: Randomization strategies for iterative data mining. In Proc 15th ACM SIGKDD, pages 379-388, 2009. http://dx.doi.org/10.1145/1557019.1557065

Short biography:
Dr Kai Puolamäki (http://www.iki.fi/kaip/)is Senior Research Scientist at the Finnish Institute of Occupational Health. He completed his PhD in 2001 in theoretical physics at the University of Helsinki. Dr Puolamäki holds a title of docent in information and computer science at the School of Science of Aalto University, Finland. His primary interests lie in the areas of data mining, machine learning, and related algorithms.