Informal seminar series

Our informal seminar series runs fortnightly on Thursdays at 2pm. Please join the ADSC mailing list to receive information on any upcoming events, and contact datascience@adelaide.edu.au if you would like to present.

Below are examples of our informal lunchbox talks, where data science experts discussed cutting-edge research findings.

A Multivariate Auto-Logistic Actor Attribute model

Dr Koskinen provided the following abstract:

Graphical models provide a modelling framework for dependent binary variables derived from principled dependence assumptions. However, these log-linear models offer little guidance for modelling binary variables for non-independent observations. Dr Johan Koskinen proposes a class of models that simultaneously account for dependencies among variables and between observations. In particular, the research aims to provide a model for vectors of binary variables, where outcomes may be subject to social influence through an exogenously define network on the observation units. The model defines a set of specific dependence assumptions that in turn define a dependence graph on the outcome variables. The implied log-linear model then turns out to be a conditional exponential family random graph model on a one-mode plus two-mode network, conditionally on the one-mode network amongst observation units. As such, this model generalises the auto-logistic actor attribute model (Robins, Elliot, and Pattison 2001; Daraganova, 2009; Koskinen and Daraganova, 2022) for univariate binary outcomes. In this talk, the guest speaker demonstrates how relaxing some sets of assumptions yield log-linear models with different interactions and, in particular, how the standard graphical model for independent observations becomes a special case. The ultimate aim, then, will be to determine how much of the dependence among variables is explained by dependence among observations. The discussion provides some initial insights and the application of a simple model to a dataset on the uptake of modern contraceptive techniques in rural Kenya.

This is joint work with Peng Wang, Neelam Modi, Jonathan Januar, and Noshir Contractor.
 

Teaching data science

Dr McIver provided the following abstract:

When we think of teaching data science, we typically think about teaching statistics, programming, particular modules or software, and probably some machine learning as well. It’s awfully tempting to teach data science as a set of cookie cutter processes, divorced from the context. We often forget the foundational need for what I call rational data skepticism. Teaching data science in high schools is even more difficult because not only are teachers largely unfamiliar with data science but they are actually often afraid of it. 

It turns out that teaching data science using authentic data sets and real problems helps us solve both of these issues. The beauty of using real, rich, complex datasets to teach data science is that you can’t eyeball the data and guess the average. You can’t look up the result in the back of the book. You have to critically evaluate your own work, starting with the assumption that there are flaws in the data, and flaws in your results, and you can build up data analysis skills by starting with basic spreadsheet skills and gradually increasing the complexity. 

Come with me and explore the use of data science as a motivator to teach tech, critical thinking, and effective problem solving. 

Complex network models for the spread of social contagion

Dr Ye provided the following abstract:

Social contagion refers to behaviours, information, rumours and opinions that spread through social networks. The spreading mechanism is due to person-to-person interactions (micro-dynamics), and often leads to emergent collective population phenomena (macro-dynamics). Applications include the outbreak of protests and riots, the adoption of new conventions and norms, and spread of misinformation. This talk will focus on the use of complex network models to study the spread of social contagion.
 
In the first part of this talk, Dr Ye will introduce a standard complex network model for studying social contagion, grounded in evolutionary game theory. He will discuss how we used data from group experiments to refine key features of the model, leading to more realistic predictions and insight on when social tipping points may occur. The second half of the talk will explore the use of complex network models to study different problems involving social contagion, including deriving theoretical guarantees for observing social tipping points, computational methods for identifying key individuals in a network to act as agents of change, and how information and conventions have fundamentally different spreading mechanics.