You can Find the schedule and abstracts below. The appropriate Zoom link can be found in the schedule.

3:30:00 PM

Achini Erandi - The University of Melbourne

Finding minimum staffing requirement of a blood donor centre

Australian Red Cross Lifeblood collects blood from almost entirely non-remunerated voluntary donors. Therefore, it aims to improve donor satisfaction by reducing waiting times and optimising staff hours at the same time. Aligning donor arrivals, staff capacity, and shifts is a key step towards reducing waiting times. We start by developing a simulation model that captures almost all of the uncertainty in the donation process and use it to compute the average waiting time. Our objective is to implement a method for determining the optimal staff roster based on the predicted staffing demand via two phases. First, we establish minimum staffing requirements to ensure that the system's predicted average waiting time does not exceed a certain limit. In the second phase, we find an optimal staff roster that meets the minimum staffing requirements. I shall present how we apply the simulated annealing algorithm to find the minimum staffing requirements and sensitivity analysis to determine the key parameters.

3:50:00 PM

Fan Cheng - Monash University

Manifold learning with approximate nearest neighbors

Analyzing high-dimensional with manifold learning algorithms often requires searching for the nearest neighbors of all observations. This presents a computational bottleneck when data size is large or when observations lie in more general metric spaces, such as statistical manifolds. We resolve this problem by proposing a broad range of approximate nearest neighbor (ANN) methods to be used within manifold learning. The novelty of our evaluation of ANN algorithms is the manifold learning setting, with algorithms compared on the basis of embedding accuracy. A second novel contribution is to use ANN for statistical manifolds by exploiting the connection between Hellinger/Total variation distance for discrete distributions and the L2/L1 norm. A thorough empirical investigation of the benchmark MNIST dataset shows that ANN algorithms substantially improve computational time with little to no loss in the accuracy of manifold learning embedding. The result is robust to different manifold learning algorithms, different approximate nearest neighbor algorithms, and different measures of embedding accuracy. The proposed method is applied to learning statistical manifolds of electricity usage. This application demonstrates how underlying structures in high dimensional data, including anomalies, can be visualized and identified, in a way that is scalable to large datasets.

4:10:00 PM

Laurence Davie - Queensland University of Technology

Bayesian Detectability of Airborne Induced Polarisation using Reversible Jump Sequential Monte Carlo

Detection of induced polarisation (IP) effects in airborne electromagnetic (AEM) measurements does not yet have an established methodology. This contribution develops a Bayesian approach to the IP-detectability problem using decoupled transdimensional layered models, and applies an approach novel to geophysics whereby transdimensional proposals are used within the embarrassingly parallelisable and robust static Sequential Monte Carlo (SMC) class of algorithms for the simultaneous inference of parameters and models. Henceforth referring to this algorithm as Reversible Jump Sequential Monte Carlo (RJSMC), the statistical methodological contributions to the algorithm account for adaptivity considerations for multiple models and proposal types, especially surrounding particle impoverishment in unlikely models. Methodological contributions to solid Earth geophysics include the decoupled model approach and proposal of a statistic that use posterior model odds for IP detectability. A case study is included investigating detectability of IP effects in AEM data at a broad scale.

4:30:00 PM

Jacob Priddle - Queensland University of Technology

A Bayesian Personalised Risk Model for Liver Fluke Infection

Precise and focused risk assessments of liver fluke infection in cattle can be used to increase awareness and promote management uptake. However, accurate estimation of the risk of liver fluke infection in cattle is challenging. Liver fluke disease in adult cattle is typically subclinical -- meaning infected animals often have no visible symptoms. This is exacerbated by the complex liver fluke lifecycle, which is highly sensitive to climate conditions and requires the presence of the intermediate snail host. The aim of this work was to create a predictive modelling tool that can be used to predict the location-specific risk of infection with respect to climate conditions. The study utilised abattoir data of over 4 million cattle processed between 2016 and 2020 at Teys' abattoirs. A binary indicator variable for liver fluke infection (liver fluke or no liver fluke) was observed at processing for each animal, with no further information as to when the infection may have occurred. By viewing the data within a survival analysis framework, we propose a Bayesian proportional odds model to predict the risk of infection depending on the age and location of the animal. The model updates the estimated risk of infection according to location-specific and time-varying covariates. The model output was used to create personalised risk profiles for each post-code. The work in this study can be used to help inform farmers about the risk of infection on their property.

4:50:00 PM

Prosha Rahman - UNSW Sydney

Linear regression in the big data context

Big data samples are convenient and large, but typically suffer from coverage issues. We can "fill in" the uncovered responses through non-parametric means, thereby creating a pseudo-census. Models, and their associated estimators, can then be fitted to the pseudo-census data. In this seminar we utilise the KNN algorithm to impute the uncovered responses and show its effects in the multivariate linear regression setting.

11:30 AM

Aishwarya Bhaskaran - University of Technology Sydney

Usable and precise asymptotics for generalized linear mixed model analysis

Despite the large volume of research concerning generalized linear mixed models, there is very little theory concerning the statistical properties of maximum likelihood estimators of these models.

In this talk, an overview of how we derive precise asymptotic results that are directly usable for confidence intervals and Wald hypothesis tests for likelihood-based generalized linear mixed model analysis will be presented. The essence of our approach is to derive the exact leading term behaviour of the Fisher information matrix when both the number of groups and number of observations within each group diverge. This leads to asymptotic normality results with simple studentizable forms.

11:50 AM

Trevor Matthews - The University of Adelaide

Estimating Ambulance Drive Time

SA Ambulance Service has a multi-tiered series of ambulance resources that the Ambulance Dispatcher must allocate to incoming requests for service. My research is aiming to explore predictive modelling to optimise the selection of the right clinical resource based on historical data. Proving that this approach is "better" than the current response process obviously requires some form of analysis.

Theoretical analysis of the benefit of a proposed change in dispatch rules in an Emergency Service is complicated by multiple periodic functions with respect to case frequency, the requirement to model the travel times within the service period and the geo-spatial influences also involved. This, coupled with the risk adverse nature of the provision of these services with respect to changing established practice, may require real-world simulation (or the development of a 'sim-city') of the environment to enable comparison between the current dispatch policies and the proposed changes and resultant cost/benefit analysis in a manner acceptable to the service provider.

Simulation of real-world movement of resources in this simulated environment is not as simple as using online mapping tools. One example of the difficulty is that the online estimates of travel time are not realistic compared to observed, real world, emergency (lights and sirens) driving. This talk will discuss the operational research of interest to the researcher, and a model developed to satisfy the requirement for a reasonable estimation of response time under emergency driving rules.

12:10 PM

Virginia He - University of Technology Sydney

Bayesian Generalized Additive Models Selection Including a Fast Variational Option

Large and complex data sets will continue to grow due to ongoing technological advancements. For flexible and interpretable regression models for such data, a problem of paramount importance is choosing among the candidate predictors and categorising their effects to be zero, linear or non-linear. Using paradigms such as spike-and-slab (group) least absolute shrinkage and selection operator priors, Markov chain Monte Carlo and mean field variational Bayes we derive scalable algorithms for reliable three-category generalized additive model selection. An R package named gamselBayes accompanies our methodology.

12:30 AM

Vektor Dewanto - The University of Queensland

Reinforcement learning from transient states: a discounting-free approach

Typical reinforcement learning works with a discount factor, and for environments whose states are all recurrent (being visited infinitely often). We aim to refine such a tradition by proposing a new policy gradient method. It is capable of finding a policy that is nearly optimal not only in recurrent, but also transient states. It is also discounting-free as it optimizes Veinott's objective (which is originated in dynamics programming). Experimental results provide insights into the fundamental mechanisms of our proposal.

12:50 pm

Punya Alahakoon - The University of Melbourne

Maritime and quarantine in Australia during the flu pandemic of 1918-19: perspectives based on mathematical modelling

The influenza pandemic of 1918-19 was one of the most devastating pandemics of the 20th century. It killed an estimated 50-100 million people worldwide. In Australia, the death toll was estimated to be 15 000.

In late 1918, when the severity of the disease was apparent, as the first line of defence to try and prevent the virus from reaching Australia, the Australian Quarantine Service was established. The vessels that had travelled overseas and inter-state were intercepted and people were examined and quarantined for a few days. Some of these vessels were carrying the infection throughout their voyage and cases were prevalent even by the time the ship arrived at a Quarantine Station. I focus on records submitted to the Quarantine Service by medical officers on board and Quarantine staff regarding the influenza outbreaks. In this talk, I will show how data pre-processing, mathematical modelling, and statistical inference can be used to understand the dynamics of influenza in contained environments such as ships.