Researcher

Dr Clara Grazian

My Expertise

  • Bayesian statistics
  • Statistics for genomics
  • Effect of air pollution on environment and human health
  • Effect of human activities on animal behaviours

Keywords

Field of Research (FoR)

Biography

Clara Grazian received a joint PhD in 2016 from Université Paris-Dauphine under the supervision of Prof. Christian Robert and from Sapienza Università di Roma under the supervision of Prof. Brunero Liseo.

Her research interests include: Bayesian statistics, mixture models, spatio-temporal modelling and copula models and variable selection, with applications in climatology, epidemiology, psychology and genetics.

Before joining UNSW, she was...view more

Clara Grazian received a joint PhD in 2016 from Université Paris-Dauphine under the supervision of Prof. Christian Robert and from Sapienza Università di Roma under the supervision of Prof. Brunero Liseo.

Her research interests include: Bayesian statistics, mixture models, spatio-temporal modelling and copula models and variable selection, with applications in climatology, epidemiology, psychology and genetics.

Before joining UNSW, she was Postdoctoral Fellow at the University of Oxford, working on understanding genomic mechanisms conferring resistance to tuberculosis and at the Università degli Studi "Gabriele d'Annunzio" working on understanding the effect of pollution on human health.
 


My Grants

October 2020 - March 2021: ACEMS Industry Collaboration Support Scheme (ICSS), Research Project: "Developing Bayesian methods for modelling the dynamics of complex systems in sports: perfomance vs injuries" (20,000$), in collaboration with QUT. 

October 2020 - September 2021: Defence Innovation Network Scheme, Research Project: "Optimising ADF military working dog performance through next-generation monitoring systems" (175,000$), in collaboration with University of Sydney and Western Sydney University.

July 2020 - June 2021: Defence Science Partnering Deed - Schedule Research Agreement (Defence Science & Technology Group - DSTG), Research Project: "Knowledge Sythesis for Autonomous Analyst" (160,000$), in collaboration with UTS and University of Wollongong.

 

 

 


My Qualifications

PhD in Mathematical Statistics (2016) - Université Paris-Dauphine and Sapienza Università di Roma

Master 2 in Mathematical Statistics (2011) - Université Paris-Dauphine

Master in Statistical Sciences (2012) - Sapienza Università di Roma

Bachelor in Statistical Sciences (2012) - Università degli Studi di Torino


My Awards

Here are the received grants since 2015

2018 - Research Project of National Interest (Italian Ministry of Research and Education) - "Environmental processes and human activities: capturing their interactions via statistical methods", (principal investigator: Daniela Cocchi)

2017 - International Exchanges Scheme, The Royal Society - "High-dimensional Bayesian dependence modelling with conditional copulas" (with Luciana Dalla Valle, Julian Stander and Brunero Liseo)

2016 - Research Project of National Interest (Italian Ministry of Research and Education) - "Likelihood-free methods for inference" (principal investigator: Brunero Liseo)

2015 - International Exchanges Scheme, The Royal Society - "Empirical and Bootstrap Likelihood procedures for Approximate Bayesian Inference" (Fabrizio Leisen and Brunero Liseo)

 

 


My Research Activities

Dr Grazian is interested in Bayesian analysis for complex systems, in particular for clustering and dependence modelling. Her research is focused on developing consistent techniques for Bayesian clustering in several situations (independent data, temporal data, spatial data) and evaluating dependence functionals in a semi-parametric setting. She works in several applied settings, in particular in genomics, finance and environmental sciences:

  • clustering approaches to identify several levels of resistance in M. Tuberculosis
  • spatial clustering for environmental variables
  • temporal clustering to model animal behaviour
  • functionals of dependence for log-returns 
  • conditional functionals of dependence for building materials

My Research Supervision


Areas of supervision

  • Mixture models
  • Copula models
  • Bayesian spatial modelling
  • Clustering
  • Spatio-temporal modelling
  • Applications in genomics, cyber-security, ecology, finance

Examples of project

1.  Data-driven approaches for determining the rate of somatic mutation in individual cells

In collaboration with the Garvan Institute of Medical Research (Dr. Manu Singh, Prof. Tim Peters, A/Prof. Fabio Luciani): 

We aim to estimate the rate of somatic mutation along a clonal evolutionary tree from single-cell genomic sequencing. The project is analogous to estimating the value of parameter K from a Dirichlet multinomial distribution (i.e. estimating the total number of unique ball colours in a Polya urn), with two added variables: i) The urn draws are not true replicates in that each one is situated in phylogenetic space along an evolutionary tree, and ii) there is assumed to be a degree of observational error in the draws, incurred by incorrect genotyping from the sequencer.

The candidate will be expected to deploy and test a number of existing computational approaches, with the potential of developing novel algorithms and publication in high-impact journals. They will have a strong background in computer science and/or software engineering, and be comfortable operating in Linux and high-performance computing environments. A background in mathematics and/or statistics, especially probability distributions, will be looked upon favourably.

2. Characterising animal behaviour through spatio-temporal modelling

Judgement bias is a promising tool for objectively measuring positive and negative affective states in animals. It is unknown whether an individual’s behaviour could be used to predict their judgement bias. Some candidates include displacement behaviours, but displacement behaviours in animals may also be displayed when animals are transitioning from one type of activity to another. Are there behavioural indicators of optimism or pessimism in domestic dogs? Are there ways to quantify fatigue in dogs and highlight the need for rest? We use temporal modelling for characterising animal behaviours and being able to cluster groups of behaviours in classes in order to identify change points. 

3. Approximate Bayesian computation for kinetic models 

In collaboration with the Sydney School of Health Science

Dynamic Positron emission tomography (PET) is a non-invasice tomographic imaging technique that produces quantitative 3D maps of physiological parameters. The mapping from the images to the physiological parameters usually requires time-series analysis of a dynamic sequence of images using kinetic models and optimisation algorithm. We use a Bayesian approximate method to derive posterior distributions of important parameters which can be implemented in high-dimensional images. 

4. Copula representation of the Black and Litterman model for portfolio optimisation

The Black and Litterman model provides an easy way to combine investors' views and market information to obtain the definition of a portfolio. More recently, Meucci (2005) introduced a portfolio reallocation method based on a copula representation of the Black-Litterman model which relax the Gaussian assumption of the financial time-series and is more robust to extreme events. We want to work in the definition of a Bayesian alternative to this representation which probabilistically optimise the portfolio. 

5. Analysis of long-memory financial sequences

Long memory is a phenomenon which may arise in stochastic processes, such that the correlations and the variance of the sample mean decays at a slower rate than 1/n. A characterising property of long-memory processes is that the absolute autocorrelations are not summable. The definition of a likelihood function for the long memory parameter is still limited, therefore standard testing method, to assess if a process has long-memory, are limited. We propose an approximate Bayesian method to derive a posterior distribution for the long-memory parameter. 

6. Spatio-temporal clustering methods based on logit-N distribution

Spatio-temporal clustering based on mixture models presents an intrinsic difficulty due to the fact that usually latent Gaussian distributions are assumed to model spatio-temporal dependence, while the probabilities to belong to each of the clusters define a compositional vector, i.e. a vector of values summing up to a constant. Such constrain introduces a particular dependence structure among the probabilities which may modify the dependence structure imposed with the Gaussian process. Such difficulty influences the spatio-temporal modelling, in particular when clustering is the goal of the analysis. For example, in environmental science, it is important to cluster observations characterised by similar levels of temperatures, rainfalls or pollution. We propose a method to overcome this difficulty. 

 


Currently supervising

  • Louis Ye (Honours)
  • Adam Stanley (Honours)
  • Yuan Gong (Master)
  • Gan Xu (Master)
  • Tong Wang (Master)
  • Vasiliki Vamvaka (Master)

Past supervision: 

  • Diego Battagliese (PhD in Mathematical Methods for Economics and Finance, Sapienza University of Rome): "Penalising Model Complexity" (joint with Prof. B. Liseo)
  • Yiwei Wang (Master of Statistics): "The invertibility of biometric representations through CNN"
  • Taiming Xu (Master of Financial Mathematics): "Copula representation of the Black and Litterman model" 
  • Xinyu Guan (Master of Statistics): "Evaluating M. Tuberculosis resistance through mixture models"
  • Wei Li (Master of Statistics, University of Oxford): "Establishing antimicrobial resistance breakpoints via mixture models"
  • Ilaria Masiani (Master of Statistics, University Paris-Dauphine): "On proper scoring rules for Bayesian model selection" (joint with Prof. C. P. Robert)
  • Alex McInnes (Bachelor in Science): "An application of copulas to OPEC's changing influence on fossil fuel prices"

My Teaching

View less