Announcing a meeting of the Statistical Society of Australia, W.A. Branch.
6:00 ᴘᴍ (AWST) on Tuesday 12th May 2020
Zoom Online Meeting (Join from 5:45 ᴘᴍ for general socialising)
Young Statisticians Meeting
Presenting work undertaken at Curtin University
Presenting work undertaken at Murdoch University
Restructuring categorical predictors using L1-norm penalties, with an application to stillbirth prediction
The LASSO (Least Angle Selection and Shrinkage Operator) is widely used for constructing sparse regression models. It is a penalised regression method that incorporates an L1-norm penalty to enforce selection and shrinkage of coefficients. However, sparse modelling using only categorical predictors is especially challenging, for two reasons:
- The number of parameters that require estimation can be very large when categorical variables are re-encoded as indicator variables, and near collinearities can occur; and
- Selecting indicator variables without regard to the underlying structure of the levels within a predictor yields models that are difficult to interpret.
Hence, alternative penalties are required that yield sparse models which respect the underlying structure of a predictor variable, such as levels that have a similar effect on the response variable.
Tutz and Gertheiss (2010) have introduced a pair of structured fusion penalties, one for ordinal predictors, the second for nominal predictors, that penalise pairwise differences of coefficient estimates. These penalties provide an alternative to the LASSO and group LASSO that yield interpretable sparse regression models when the predictors are categorical variables. This talk will introduce geometric interpretations of each penalty and illustrate their use by constructing sparse logistic regression models for predicting stillbirth from categorical variables. The results show that using these structured fusion penalties yields models that have a sensible interpretation at no cost to predictive ability compared to using unstructured penalties.
ABOUT THE SPEAKER:
Emily Whitney graduated from Curtin University mid-2019 with 1st Class Honours in Mathematical Sciences. She was awarded the society's 2019 grant which she used to attend the International Workshop on Statistical Modelling in July 2019. Her work on predictive models for stillbirth inspired a passion for all things 'public health', leading her to recently beginning a consultant role where she currently works as a data scientist in the health informatics space.
A comparison of automated procedures for robust multivariate outlier detection
The rise of automatic outlier detection procedures remains increasingly important particularly in the field of multivariate analysis. This comes at a time of the technological era which has facilitated the demand for new and robust procedures that are able to handle the large supply of data that has become synonymous with it. Although recent research in the field has focused on the implementation of machine learning techniques there remains a demand for simple approaches that are far less computationally intensive. Three prominent outlier detection algorithms will be explored with comparisons made through simulation of multivariate data generated from both contaminated and uncontaminated distributions over a wide range of imposed conditions. A particular motivation for this study is the focus on a specific adaptive method known as the adaptive trimmed likelihood algorithm (ATLA).
ABOUT THE SPEAKER:
Andrew Grose recently graduated from Murdoch University where he completed an Honours in Mathematics and Statistics under the supervision of Dr. Brenton Clarke. The topic of his Honours thesis, which also serves as the basis of this talk, involves an exploration into Outlier detection and in particular how this interfaces with multivariate data. Andrew has since started working at SAGI-West, Curtin University as a Biometrician.
This seminar will be presented online using Zoom. To receive the connection instructions please register on this page. Instructions will be sent to your email on registration. There is no close-off time for registrations though it is recommended to register in advance.
Once the seminar begins, participates will be asked to mute themselves. The meeting will be interactive, and viewers will be able to ask questions.
For further information please contact the Branch Secretary, Rick Tankard, Murdoch University.
He can be reached by email at email@example.com or by phone at (08) 9360 2820.