Schedule

9:30 am - Pre-conference workshop (420min)
See the workshop page for details.
6:00 pm - Social event (180min)
Pre-conference informal gathering	📍The Lighthouse Pub, Belconnen

8:00 am - Registration desk open (60min)

9:00 am - Conference open (60min)
Welcome to Country	Ngunnawal Elder Uncle Wally Bell
Splinkage and Other Stories: Data Science at the ABS Abstract National statistical offices around the world have gained a great deal from open-source data science, and the Australian Bureau of Statistics is no exception. In this address, I will tell the story behind two recent ABS innovations, joint winners at this year’s APS Data Awards. The first is our adoption of Splink, an open-source probabilistic linkage tool, to build the Person Linkage Spine that underpins the Person-Level Integrated Data Asset (PLIDA). Linking the Spine datasets now takes 4 business days rather than 54, and we have contributed a new clustering method back to Splink. The second is the ABS Data Coding Capability, a hierarchical machine learning service that codes free-text occupation responses to the new Occupation Standard Classification for Australia (OSCA) and is now coding the 2026 Census. I will finish by reflecting on what both stories show about expertise, collaboration and trust.	Dr David Gruen AO 1. Australian Bureau of Statistics
10:00 am - Morning tea (30min)

10:30 am - Keynote (60min)
Decision-Making with Statistics and Data Science in the Age of AI Abstract Artificial intelligence (AI) is emerging as one of the most significant technological disruptors of contemporary society. It is changing almost every aspect of our lives: how we generate and discuss ideas, gather and process information, automate workflows, write code, solve mathematical problems, and ultimately make decisions. In this talk, I will present some emerging uses of AI in statistics and data science. I will also reflect on trends I am increasingly seeing as the Editor-in-Chief of the R Journal, which publishes research centred on R, one of the most widely used programming languages for statistics and data science. From this perspective, I have witnessed first-hand how AI is changing journal article writing, R package development, peer review, and copy-editing. I will discuss both the risks and opportunities that AI brings. As a society, we need to recognise and embrace the distinctive value that humans bring to the table, while adapting to new ways of working, reasoning, and making decisions alongside AI.	Dr Emi Tanaka¹ 1. Australian National University
11:30 am - Panel session (60min)
Who’s Doing the Analysis? Data Science in the Age of AI Agents Abstract AI has moved rapidly from answering questions and generating code to acting as an agent: exploring data, modifying code, running analyses, using external tools and working through complex tasks with increasingly little human intervention. For data scientists, this raises questions that go well beyond whether an LLM can write a useful piece of code. How much of an analytical workflow should we delegate? How do we verify work produced by an AI agent? Does supervising an agent save time, or simply create a new kind of work? And what expertise do data scientists need when producing an analysis becomes easier than knowing whether it is any good? This panel brings together perspectives from across data science to discuss where LLMs and AI agents are genuinely useful today, where they still fail, and how they may change the practice of data science. We will explore trust, reproducibility, human oversight, accountability, changing skills, and what happens when the data scientist increasingly becomes the manager, rather than the analyst.
12:30 pm - Lunch (60min)
ADSN into the future (12:45pm to 1:15pm) Abstract Grab some lunch and come for a chat about what’s happened in the past year and where the ADSN is heading next. We want to know what has been valuable to you so far if there are any new initiatives you’d like to see.	Ben Harrap
1:30 pm - Invited talks: Decision Support (90min)
A Data-Driven Framework for Quantifying Airspace Risk Across Australia Abstract Turning heterogeneous spatial and temporal data into defensible decisions is a challenge across many sectors. The Australian Digital Airspace Characterisation (ADAC) project, led by QUT with Boeing and FlyFreely, shows how a national-scale, data-driven framework can turn billions of observations into transparent, quantitative insights informing operational planning and policy. ADAC evaluates over 800 billion proximity events across more than 200 billion locations spanning Australia and its coastlines, integrating satellite and terrestrial sources with aeronautical, terrain, obstacle, population and land-use data. Through advanced statistical and probabilistic modelling, it produces continuous, high-resolution estimates rendered as interactive national-scale maps. The presentation covers ADAC’s methods and decision-support applications: computational architecture for continental-scale modelling, rare-event methods for quantifying uncertainty in sparse-observation environments, and visualisation turning outputs into actionable maps. We illustrate impact through case studies in health logistics, emergency services, and resource operations, and note stakeholder engagement with CASA.	Aaron McFadyen Guilherme Froes Silva Peter Smith Jason Brown
From Principles to Decisions: An End-to-End GenAI Evaluation Framework Abstract Government agencies are increasingly adopting Generative AI(GenAI) to improve public services and information dissemination. However, the variable and context-dependent behaviour of large language models makes it difficult to determine whether these systems are safe, reliable and ready for deployment.While Responsible AI principles provide high-level guidance, practical methods for operationalising them remain limited across the public sector. This work presents a reusable, end-to-end GenAI Evaluation Framework that supports responsible deployment decisions.The framework combines domain-driven scenario design, automated generation of evaluation datasets, and LLM-as-Judge scoring to assess quality, performance and safety across measures including relevance, grounding, clarity, accuracy, tone, confidentiality and toxicity. Using a Census chatbot as a case study, we demonstrate how the framework transforms evaluation from an ad hoc activity into a continuous, structured and evidence-based process. By bridging Responsible AI principles and engineering practice, it provides a scalable approach for assessing GenAI systems and supporting trustworthy deployment decisions across government.	Najmeh Samadiani¹ Ilana Lichtenstein¹ 1. Australian Bureau of Statistics, Melbourne, Australia
Artificial Intelligence for Medical Imaging: Enhancing Diagnostic Accuracy Through Deep Learning Abstract Medical image analysis is increasingly benefiting from artificial intelligence (AI), offering opportunities to improve diagnostic accuracy and clinical decision-making. This research investigates deep learning approaches for automated classification of medical X-ray images, focusing on musculoskeletal and chest radiographs. The study combines transfer learning, feature fusion, and attention-based convolutional neural networks to enhance feature extraction and classification performance. Semi-supervised learning techniques were also explored to address limited labelled data. The proposed frameworks were evaluated on multiple datasets, including abnormality detection in musculoskeletal X-rays and classification of chest X-rays containing COVID-19, pneumonia, tuberculosis, and normal cases. Results demonstrated high classification accuracy and improved model robustness. Explainable AI techniques were applied to visualise regions influencing model predictions, supporting transparency and clinical trust. The findings demonstrate the potential of AI-assisted diagnostic systems to improve healthcare outcomes and support future clinical implementation.	Zaenab Alammar^1,2 Laith Alzubaidi^2,3 Jinglan Zhang^1,2 1. School of Computer Science, Queensland University of Technology, Brisbane, Australia 2. Centre for Data Science, Queensland University of Technology, Brisbane, Australia 3. School of Mechanical, Medical, and Process Engineering, Queensland University of Technology, Brisbane, Australia
Bridging Digital Health Innovation and Data Governance: Lessons from Accessing Statewide Electronic Medical Records for Secondary Research Abstract Digital health systems generate vast clinical data, creating opportunities for evidence-based policy, planning, and care. Queensland’s integrated electronic Medical Record (ieMR) system offers a valuable platform for population-scale research and secondary data use. This presentation explores the pathway from data availability to decision-making through a case study of accessing de-identified ieMR data for research. Despite mature infrastructure and established partnerships, approvals took nearly two years due to fragmented ethics governance, lengthy Public Health Act processes, and concerns about data scale. The experience shows that evidence generation depends not only on analytical capability but also on governance, stakeholder trust, and social licence. Beyond legal compliance, researchers must demonstrate public benefit and responsible data stewardship. Streamlined governance, transparent engagement, and trusted access frameworks are essential to bridging the gap between technical capability and practical data use, enabling the full value of health data to support research, policy, and innovation	Danyang Dai¹ Pedro Franca Gois^2,3,4 Sally Shrapnel⁵ SMART Hub¹ Jason D. Pole¹ 1. Queensland Digital Health Centre, The University of Queensland, Brisbane, Australia 2. Faculty of Health, Medicine & Behavioural Sciences, University of Queensland, Brisbane, Australia 3. Nephrology and Transplantation, John Hunter Hospital, Newcastle, NSW, Australia 4. School of Biomedical Sciences and Pharmacy, University of Newcastle, Newcastle, NSW, Australia 5. School of Mathematics and Physics, University of Queensland, Queensland, Australia
Multi-Model Ensemble for Automated Longitudinal Electronic Health Records Summarisation in General Practice Abstract General practice (GP) clinicians managing multimorbidity face burden from fragmented records. Automated longitudinal EHR summarization is underexplored. This study introduces an AI pipeline to evaluate accuracy, completeness, and utility. We adapted ACSQHC Summary guidelines for GP. The pipeline integrates rule-based extraction with LLM ensembles coordinated by a validation agent enforcing schema constraints. Deployed locally on 2,232 multimorbidity patients from GPs in South West Sydney, two clinicians reviewed 500 for error and utility via adjudication. Single-model ablation variants were benchmarked against metrics. All 2,232 records achieved 99.88% validity. For 1,500 instances: inaccuracy 2.8%, hallucination 9.7%, omission 10.5%. Omission peaked in DIAGNOSES (15.0%), hallucination in CLINICAL SUMMARY (12.1%). Utility rated 4/5 by reviewers. Ensemble designs proved more balanced than single-model variants. Automated EHR summarization in GP is feasible with local LLM ensembles but mitigating omission and hallucination remains a challenge. This study establishes a reproducible framework, demonstrating tangible improvements in multimorbidity care via adapted guidelines.	Vu Thinh Doan^1,2, Bocheng Li^3,4, Jiayue Guo⁵, Hsieh-Chih Hsu¹, Hong-Jie Dai^1,6, Jitendra Jonnagaddala^7,8 1. Intelligent System Laboratory, Department of Electrical Engineering, College of Electrical Engineering and Computer Science, National Kaohsiung University of Science and Technology, Kaohsiung, Taiwan 2. Nha Trang University, KhanhHoa, Vietnam 3. WHO Family of International Classification Collaborating Centre, Beijing, China 4. Department of Medical Record, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China 5. School of Health Policy and Management, Chinese Academy of Medical Sciences & Peking Union Medical College 6. National Institute of Cancer Research, National Health Research Institutes, Tainan, Taiwan 7. Discipline of General Practice, School of Clinical Medicine, UNSW Sydney, Kensington, Australia 8. SREDH Consortium, Sydney, Australia
3:00 pm - Afternoon tea (30min)

3:30 pm - Invited talks (30min)
Sciget.org: Developing a national Virtual Research Environment for data analytics Abstract Australian researchers contend with fragmented AI/ML infrastructure, limited GPU access, incompatible software installations, and no consistent governance for sensitive data. Here we present the design and first delivery phase of a national Virtual Research Environment (VRE): a portable, containerised platform delivering advanced analytics and AI capabilities. The Sciget.org VRE combines an automated container build system, a federated software distribution network based on CERN Virtual Machine File System (CERN VM-FS) (a web-based, global versioning file system optimised for software distribution), GPU-enabled compute, and a virtual desktop. Privacy-preserving large language models offer in-environment coding support. A three-tier architecture matches security to data sensitivity: open environments for public datasets and early-career training, application-level ISO 27001-aligned environments for moderately sensitive health data, and a foundation for future Trusted Research Environments. We demonstrate an exemplar VREs for the Australian microscopy and the AI/ML community supporting data science and advanced analytics workflows.	Steffen Bollmann¹ Aswin Narayanan¹ Glen Charlton² Long Le² Gnana K Bharathy³ Aleem Uddin³ Elleina Filippi³ Celia Loneragan³ Mona Bojamma³ Jayadeep Jacob³ 1. University of Queensland 2. Intersect Australia 3. Australian Research Data Commons
Advanced Analytics and AI Resources Hub: Enabling Data-Driven Decisions in Health Research Abstract In 2023, the Australian Research Data Commons (ARDC) launched a national digital research infrastructure initiative to strengthen Australia’s capacity for digital health research and translation. As part of this program, an ARDC-ADSN collaboration undertook a comprehensive assessment of current digital health infrastructure, including computing capability, data availability, analytics methods, and accessibility. This work identified critical national research infrastructure (NRI) requirements needed to support advanced health data analytics and informed the design of a national framework. Building on insights from Phase 1, which engaged researchers, government stakeholders and industry partners, Phase 2 applied a structured, collaborative, co-design methodology to develop a sustainable Resource Hub aligned with national health research priorities and the ARDC Advanced Analytics Reference Architecture. The approach was organised around five thematic areas, each addressing a key component of the digital infrastructure stack. This presentation will provide an overview of the project and summarise the major findings from this work.	Divya Mehta¹ Bernadette Hyland-Wood² Alberto Nettel Aguirre³ Jessie Roberts¹ Michael Guihot¹ Catarina Pinto Moreira⁴ Dimitri Perrin¹ Olivier Salvado¹ Kerrie Mengersen¹ Gnana Bharathy⁵ Nicola Armstrong³ 1. Queensland University of Technology 2. Curtin University 3. University of Technology Sydney 4. Australian Research Data Commons
4:00 pm - Fast-forward presentations (15min)
How to make a good variable naming convention: guidelines for development and application in research data contexts Abstract Guidelines for good data management and reproducible research practices routinely list data dictionaries as an essential document to accompany datasets, with variable names being a core component of any data dictionary. Despite variable names’ unequivocal inclusion in data management guidelines, there is surprisingly little literature that comprehensively explores what makes for a good, or bad, variable name or naming convention. We present a set of guidelines for developing and applying variable naming conventions, the development of which arose from the redevelopment of an existing variable naming convention used for the Mayi Kuwayu: The National Longitudinal Study of Aboriginal and Torres Strait Islander Wellbeing. These guidelines are for researchers, data managers, and anyone else who finds themselves having to create variable names. Our ultimate goal was to produce variable names that were easy to work (code) with, easy to read and distinguish between, and robust to changes in questionnaire design.	Benjamin Harrap¹ Jonathon Martin¹ Katherine Thurber¹ Ryan Perry² Christopher McKay¹ 1. Yardhura Walani, Australian National University 2. Australian Data Archive, Australian National University
Bayesian Variable Selection in Quantile Regression to Assess Joint Impacts of Climate Indices on Monthly Rainfall Abstract Understanding how large-scale climate indices influence rainfall variability is important for assessing drought and flood risk in eastern Australia. This study examines the joint impact of the Southern Oscillation Index (SOI), Dipole Mode Index (DMI), and monthly rainfall quantiles in New South Wales (NSW), Australia, with emphasis on the tails of the rainfall distribution. To address this, we employ a hierarchical model incorporating Bayesian variable selection within a spatial quantile regression framework, allowing integration of two indices and spatial variation in impacts. The first level expresses the asymmetric Laplace distribution as a location-scale mixture. The second level uses a prior for regression coefficients as a two-component mixture, with a point mass at zero and a normal distribution. The third level models joint inclusion probabilities of SOI and DMI using multinomial logistic regression with a spatial Gaussian Process prior. Results show that SOI is the dominant driver across NSW, with impacts in winter and spring. DMI influences spring rainfall, particularly September and October. Lower and upper quantiles are more sensitive to SOI and DMI than the median.	Dilani Kaveendri^1,2 Sally Cripps³ Nandini Ramesh⁴ 1. Canterbury Institute of Management, Australia 2. The University of Sydney 3. University of Technology Sydney 4. CSIRO data61, Westmead
Evaluating Grounding Approaches for Large Language Models on Tabular Statistical Data Abstract Large language models (LLMs) offer natural-language access to complex statistical products, potentially broadening reach to high-quality data—but grounding them reliably in tabular data is hard. We compare two approaches on a curated subset of ABS demographic and socioeconomic data across standard geographies (SA2–SA4, LGA): conventional Retrieval-Augmented Generation (RAG) and Knowledge Graph-RAG (KG-RAG), using the same generation model. Measured on accuracy and coverage, each excels on different question types. KG-RAG is more accurate on direct lookup, ranking, and set-membership tasks, giving traceable, verifiable retrieval. RAG’s semantic retrieval better handles varied phrasing and achieves broader coverage on regional comparisons and area descriptions. This motivates hybrid architectures that route each question to the most suitable system. We also detail curation challenges: source metadata built for human interpretation needed transformation and normalisation before either approach could use it—underlining that effective grounding depends not only on architecture but on AI-ready data.	Ali Behnaz¹ Ivan Welsh^1,2 Bernadette Giuffrida¹ Phil Juricev-Martincev¹ Alex Tan³ 1. Australian Bureau of Statistics 2. Stats NZ 3. Baseten
A Bayesian spatio-temporal nearest neighbor Gaussian process model for pooled genetic data Abstract Large scale genetic datasets often aggregate the total allele counts of distinct genetic markers. Inferring haplotype frequencies (i.e. the frequency of multimarker alleles) from these pooled data is a challenge. Previous spatio-temporal modeling in this context has been limited to 3 markers due to the computational cost. In this work, we propose a nearest neighbor Gaussian process (NNGP) model to improve scaling with the number of markers and observations. To infer the parameters of our model, we develop a novel sequential Monte Carlo squared algorithm, which uses particle Gibbs with ancestor sampling to mutate the NNGP function values. The latter has a linear cost in the number of observations and the number of NNGPs, and can be applied to a broad range of NNGP models. As a case study, we analyze genetic data relating to antimalarial drug resistance in Africa, and show our scaling results empirically on a 3 and 6 genetic marker dataset.	Imke Botha¹ Tianxiao Hao² Lucinda E. Harrison³ Nick Golding² Daniel J. Weiss² Jennifer A. Flegg¹ 1. University of Melbourne 2. The Kids Research Institute of Australia 3. WorldWide Antimalarial Resistance Network (WWARN)
STORM-Impute: A Scalable Framework for Missing Values in Dynamic Spatio-Temporal Big Data Abstract Introduction: Missing values in large spatio-temporal datasets are often spatially clustered, temporally persistent, and generated by processes that evolve over time. Standard methods may ignore these dependencies, producing biased estimates and poorly calibrated uncertainty. Methods: We propose STORM-Impute, a Spatio-Temporal Online Reconstruction Model combining sparse Gaussian Markov random fields and state-space temporal dynamics. Sparse neighbourhood approximation, parallel computation, and online updating support high-dimensional big data. The R package stormImpute provides estimation, diagnostics, visualisation, uncertainty quantification, and benchmarking. Results: Simulation and environmental-data studies assess reconstruction error, interval coverage, preservation of Moran’s I, temporal autocorrelation, forecasting accuracy, memory use, and runtime against MICE, k-nearest neighbours, missForest, matrix completion, and spatio-temporal kriging. Conclusion: STORM-Impute offers a scalable, reproducible, uncertainty-aware solution for incomplete dynamic spatial data at scale.	Hom Nath Dhungana¹ Masood Siddiqui² Radhika Kuttala³ Eric Howard⁴ 1. Charles Sturt University, Australia 2. Lucknow University, India 3. Southern Cross Institute, Parramatta, Australia 4. Macquarie University, Australia
The data hunger problem: rethinking deep learning for clinical research Abstract Machine learning holds considerable promise for clinical decision support, yet one constraint shapes what methods are appropriate: clinical datasets are small. This talk examines that constraint through a head-to-head comparison of handcrafted feature extraction and end-to-end deep learning for paediatric respiratory disease classification from tidal breathing flow signals. We evaluated feature extraction methodologies, classical classifiers, ensemble strategies, and a variety of deep learning models trained on raw signals across approximately 350 patients with independent external validation. The CNN underperformed conventional models by 15 to 20 percent. Tabular models with hand-crafted features achieved the highest external AUC of 0.760. Deep learning’s capacity for automatic feature discovery comes with data requirements most clinical studies cannot meet. Handcrafted features encode domain-relevant signal characteristics efficiently under small-sample conditions. This talk offers a practical framework for choosing between approaches based on dataset size, class imbalance, and validation requirements, grounded in a clinical problem moving to multi-site trial implementation.	Stuart Szwec¹ Gabriela Martins Costa Gomes² Adam Collinson² Joerg Mattes² 1. Hunter Medical Research Institute 2. University of Newcastle
Modular software for sequential inference: From election audits to general data Abstract Sequential inference involves analysing data dynamically as they are collected and allowing data-based decisions, such as whether to stop early. This can be substantially more efficient/powerful than fixed-sample approaches. Possible applications include clinical trials, real-time quality control and A/B testing; our motivating application is election auditing. Implementing sequential statistical procedures in software has received comparatively little attention compared to fixed-sample procedures. There will often be complex domain-specific requirements to be met, which need to be balanced against having flexible, generic code that can be used in diverse settings. This challenge is acutely evident in election auditing, where there is substantial diversity in voting systems and electoral infrastructure. To address this need, we present an open-source software framework and Python implementation for sequential hypothesis testing. While it is geared towards election audits, we build it in modular pieces that can be easily repurposed. The core functionality implements statistical methods that are agnostic to elections, including the ability to test complex composite hypotheses.	Alexander Ek¹ Philip B. Stark² Damjan Vukcevic³ 1. KU Leuven, Belgium 2. University of California, Berkeley, USA 3. Monash University, Australia
4:15 pm - Poster session (45min)
Machine Learning Approaches for Fall Prediction and Detection in Nursing Home Residents Aged ≥60 Years Using Wearable Sensors, Environmental Sensors, and Electronic Health Records: A Systematic Literature Review Abstract Falls are a major cause of harm among nursing home residents. Machine learning (ML) using wearable, environmental, and electronic health record (EHR) data may improve fall prediction and detection, but evidence of implementation beyond algorithm performance is needed. This review assessed ML approaches and translational considerations for aged-care deployment. A PRISMA-based systematic review included studies of residents aged ≥60 years using ML with wearable sensors, environmental sensors, or EHR data for fall prediction or detection. Studies were assessed for models, validation, performance, and implementation factors. Twenty-five studies were included: wearable (n=14), environmental (n=7), and EHR-based (n=4). Wearable models showed moderate to high performance but mostly used internal validation. Environmental systems performed well in controlled settings, while EHR models showed moderate prediction ability. Implementation factors were rarely reported. ML-based fall detection systems show promise, but evidence of real-world effectiveness remains limited. Future research should focus on validation, usability, workflow integration, equity, governance, and care outcomes.	Sepideh Jahandideh¹ Ebrahim Barzegari² Mina Jahandideh³ Abbas Kouzani⁴ 1. Southern Cross University 2. Kermanshah University of Medical Sciences 3. Queensland University of Technology 4. Deakin University
ggincerta: An R package for Uncertainty Visualisation with a Layered Grammar of Graphics Abstract A key challenge in spatial visualisation is the simultaneous communication of estimated values and their associated uncertainties, as visual channels are occupied by spatial structure. We developed `ggincerta`, an R package that extends the layered grammar of graphics implemented in `ggplot2`, to facilitate the visualisation of uncertainty in spatial data. `ggincerta` enables users to leverage the full flexibility of the grammar of graphics to create complex and customised visualisations through simpler and more consistent workflows in spatial and more general data settings. The package also introduces dual map, a visualisation designed to support selective attention between values and uncertainties. We demonstrate the capabilities of ggincerta through several case studies, illustrating how the package can be used to create informative and visually effective maps. By integrating uncertainty visualisation directly into the ggplot2 ecosystem and expanding the range of available uncertainty displays, ggincerta provides a flexible and powerful framework for researchers and practitioners working with spatial and other information-rich data.	Maggie Xueqi Ma¹ Weihao Li¹ Quan Vu¹ Francis Hui¹ Emi Tanaka¹ 1. Research School of Finance, Actuarial Studies and Statistics, Australian National University
Higher Education Equity Atlas Abstract What does Australia’s higher education catchment look like? Equitable access to higher education is pivotal to Australia’s social cohesion and economic prosperity; geographic patterns of university participation are an important policy question. The Australian Centre for Student Equity and Success (ACSES) partnered with the Curtin Institute for Data Science (CIDS) to develop an interactive national mapping dashboard combining Department of Education first-address enrolment data with ABS statistical regions (SA2 and SA3) to create interactive choropleth maps of university enrolments by geography, institution, equity group and enrolment type. We discuss the challenges of cleaning and harmonising the data, handling suppressed values, optimising browser performance for the thousands of high-detail geographic regions, and designing visualisations that are both accessible and intuitive for non-technical users. The resulting geospatial platform enables researchers, universities, and policymakers to explore enrolment patterns across Australia through an interface that makes complex geographic relationships immediately apparent.	David Raithel¹ Leigh Tyers² Gemma Cadby¹ Daniel Doss¹ 1. Australian Centre for Student Equity and Success (ACSES) 2. Curtin Institute for Data Science (CIDS)
Landscape of R packages for Exploring Spatio-temporal Networks Abstract Analysing spatio-temporal network data is a contemporary research problem that combines space, time, and network structure. While techniques to analyse these dimensions separately exist and are well developed, exploring them simultaneously remains an open challenge. In exploratory data analysis (EDA), analysts need to decide which dimensions matter across potentially several EDA tasks (storage, wrangling, visualising, and modelling). That said, the existing tools remain fragmented, with each task and dimension supported by a different set of tools, particularly for dynamic networks, where structure changes over time. We review the current landscape of R packages and their ability to perform EDA tasks. We find data storage is dominated by igraph and network, but both have outdated user interactions relative to tidyverse conventions. tidygraph, which implements igraph structure as two tidy data frames, is the closest to convention. Support for interactive visualisation and dynamic networks remains underdeveloped. As for modelling, a tidyverse-style interface, similar to tidymodels, would help with data exploration.	Krisanat Anukarnsakulchularp¹ Dr. Michael Lydeamore¹ Prof. Dianne Cook¹ Dr. Natali Jokanovic² 1. Monash University 2. The Alfred Hospital and Monash University
From multimodal AI fusion to Agentic AI approach for colorectal cancer survival prediction Abstract The proposed work investigates multimodal deep learning for colorectal cancer (CRC) survival prediction using whole slide images (WSIs) and clinical data. WSIs are represented by patch embeddings extracted from pathology foundation models and aggregated using multiple instance learning and hypergraph-based modelling. Clinical variables are encoded separately and fused with image features through concatenation, gated fusion, attention-based fusion, and graph-inspired architectures. The main objective is to identify the most effective fusion strategy, model configuration, and parameter settings for robust patient-level survival prediction. A key contribution is the integration of an agentic AI layer within the framework. This layer supports autonomous management of multimodal workflows, including model selection, hyperparameter exploration, experiment tracking, and comparison of fusion strategies across datasets. Experiments use patient-level cross-validation and external validation across TCGA and MCO cohorts. Performance is assessed using AUROC, C-index, Kaplan-Meier analysis, and confidence intervals. The approach aims to improve accuracy, interpretability, and generalizability.	Miljana Shulajkovska^1,2 Matej Jelenc³ Jitendra Jonnagaddala⁴ Anton Gradišek² 1. International Postgraduate School Jožef Stefan, Ljubljana, Slovenia 2. Jožef Stefan Institute, Ljubljana, Slovenia 3. Faculty of Mathematics and Physics, University of Ljubljana, Slovenia 4. School of Population Health, Faculty of Medicine and Health, Syndey, Australia
Classifying Job Advertisements into OSCA using Calibrated Semantic Similarity Abstract This paper presents a framework for classifying job advertisements into the Occupation Standard Classification for Australia (OSCA) without directly labelled training data. Using Internet Vacancy Index data coded in Australian and New Zealand Standard Classification of Occupation (ANZSCO), the approach combines correspondence mapping with calibrated semantic similarity to construct a high-confidence training dataset. A key contribution is a threshold calibration procedure that determines when similarity scores can be trusted, enabling the selection of reliable examples while filtering out uncertain cases. Using this dataset, a BERT-based model achieves strong performance, and comparison with human annotation shows that predictions align closely with expert judgement in well-defined cases, with differences concentrated in inherently ambiguous examples. The results demonstrate that accurate large-scale OSCA classification can be achieved by prioritising label quality through calibration, providing a practical solution for updating labour market data to new classification systems.	Rui Wang¹ Yin Shan¹ Christopher McLoughlin¹ Darcy Bermingham¹ 1. Nowcasting and Labour Demand, Jobs and Skills Australia
From Match Data to Recruitment Decisions: Auditing Scouting Models for Gender Bias in Women’s Football Abstract Football clubs now use statistical models to identify and recruit players, although many performance measures were developed from men’s match data. Their validity in women’s football cannot be assumed. This study asks whether models trained on men’s data produce different ratings and recruitment rankings when applied to women players, and what safeguards should govern their use. Using Hudl StatsBomb event data from 771 women’s league matches and available men’s competitions, we will train two expected-goals models with the same algorithm, predictors, preprocessing and tuning. One will use men’s shots. The other will use a women’s sample matched for size and shot context. Both will score the same held-out women’s shots. We will compare calibration, prediction error, feature effects and ranking changes by position and league. A doctrinal analysis of the EU AI Act, GDPR and sex-discrimination law will inform a reproducible decision-support protocol for clubs and regulators. In line with ‘From Data to Decisions’, the protocol will help users test model validity before recruitment, document limitations, retain human review and allow players to challenge inaccurate assessments.	Sheikh Sultan Aadil Huque¹ Sara Mann² 1. CIL, National University of Singapore 2. University of Canterbury, New Zealand
Beyond Click Counts: A Behavioural Analytics Framework for Inferring Academic Stress Risk from Learning Management System Data Abstract Introduction: Academic stress adversely affects students and their success, but LMS analytics aim to predict performance through activity measures. There are limited behavioural insights available for early academic intervention. Methods: We propose a behavioural learning analytics framework that converts LMS interaction logs into behaviour-based signatures of learning, such as procrastination, regularity, behavioural volatility, assignment preparation, and study rhythm. In conjunction with the Open University Learning Analytics Dataset (OULAD), Explainable AI is integrated with machine learning to estimate academic stress risk and also to present explanations of the behavioural factors influencing predictions. Expected Outcomes: The framework is expected to outperform traditional click-based analytics by providing interpretable behavioural profiles and identifying key behavioural indicators associated with elevated stress risk. Conclusion: By combining behavioural analytics with Explainable AI, this work converts LMS data into transparent, actionable evidence that supports academic decision-making and student wellbeing interventions.	Radhika Kuttala¹ Hom Nath Dhungana^1,2 Eric Howard^1,3 1. Southern Cross Institute, Parramatta, Australia 2. Charles Sturt University, Australia 3. Macquarie University, Australia
torch-openreml: A PyTorch-Based Framework for Flexible Linear Mixed-Effects Modelling Abstract We introduce torch-openreml, a PyTorch-based Python library for fitting linear mixed-effects models using the Average Information Restricted Maximum Likelihood (AI-REML) algorithm. The library represents covariance models as compositional computation graphs, where complex covariance structures can be constructed by combining modular components through algebraic operations. Within this framework, model parameters can be fixed, directly estimated, or estimated on a transformed scale and mapped back to the original scale, with analytical gradients provided where possible and PyTorch automatic differentiation used for more general cases. The AI-REML solver exploits common covariance structures to improve computational efficiency and benefits from PyTorch support for CPUs, GPUs, and other accelerators. torch-openreml supports a wide range of linear mixed-effects modelling applications, such as multi-environment trial analysis, spatial modelling, and genomic prediction. It enables rapid prototyping of novel covariance models without requiring manual gradient derivation, providing an extensible platform for modern mixed-effects modelling.	Weihao (Patrick) Li¹ Emi Tanaka¹ 1. Research School of Finance and Actuarial Studies, Australian National University
6:00 pm - Social event (180min)
Conference social night	📍 Generator Gallery, Belconnen Arts Centre

8:30 am - Workshop (60min)
Hands-On Workshop: Exploring Virtual Research Environments (VREs) Abstract Discover the future of collaborative research with our new Virtual Research Environments (VREs). Tied directly to the ADSN project, these containerized, cloud-based workspaces provide seamless access to computing resources, analytical tools, and secure collaboration features all in one place. Join us for a hands-on session to test out an early alpha version of AI/ML (possible to also give pointers to other VREs completed till date) running on the Nectar cloud, and help shape the final platform with your feedback! Please bring your laptop.	Australian Research Data Commons
9:45 am - Invited talks: Infrastructure (60min)
DevOps Principles as data governance for an Aboriginal legal service Abstract Legal-assistance data for an Aboriginal and Torres Strait Islander organisation must answer to two authorities at once: external accountability to government funders, and the data sovereignty of the community it serves. We present the design of a reproducible, open-source data pipeline that turns a transactional legal case-management system into statutory compliance reports. Our central argument is that computational reproducibility — and the DevOps culture of software engineering it borrows from is not merely an efficiency practice but a form of data governance in its own right. The pipeline is organised as a climb through levels of abstraction, following the DIKW progression: raw extracts (data) are transformed by business rules into curated records (information), aggregated into statutory measures (knowledge), and used to inform service and funding decisions (wisdom). A machine-readable operational specification acts as a single source of truth — documenting business rules, validating data, and making counting definitions testable rather than aspirational. Version control, automated testing, and continuous documentation operationalise governance decisions directly as code.	Aaron Willcox¹ 1. North Australian Aboriginal Justice Agency
From open-access ocean observations to environmental intelligence: enabling data-driven decisions through Australia’s Integrated Marine Observing System Abstract The Integrated Marine Observing System (IMOS) is Australia’s national ocean observing research infrastructure, delivering sustained, open access observations of the marine environment through the National Collaborative Research Infrastructure Strategy (NCRIS). Over the past two decades, IMOS has generated more than 1.5 billion measurements spanning Australia’s polar, temperate and tropical coastal and oceanic waters, creating a rich evidence base for research, industry and government. Through the Australian Ocean Data Network (AODN), IMOS provides access to more than 200 quality-controlled marine data collections, near real-time and historical ocean information, cloud-optimised datasets, and open-source Python and R tools that support scalable analytics, visualisation, integration and emerging AI applications. This talk will showcase how IMOS enables the transition from data to decisions by supporting environmental assessment, policy, and operational decision-making across climate science, oceanography, marine ecosystems, fisheries and coastal resilience. IMOS accelerates discovery, supports data-driven decisions, and unlocks broader societal, environmental and economic benefits.	Fabrice Jaine¹ Mark Rehbein¹ Michelle Heupel¹ 1. Integrated Marine Observing System, University of Tasmania, Hobart, Australia.
Museum Data Science: Collection Data Platform and Catalogues as Data Abstract This paper introduces “Plexus” the new National Museum of Australia collection data platform. This new data platform operates as part of middleware to support existing and new data and media flows for diverse ends and end-users as “nerve centre” for collection data architecture. Highlights how the Museum as a domain producer of trusted collection data products and services is moving to an operational model that accords with two of the four data mesh concepts: data ownership and data as a product made available through an API layer for internal and external consumers. Presents the Australian Institute of Anatomy (AIA) collection card catalogue as a case study. The AIA catalogue demonstrates the value of catalogues “as data” with multiple sources of data and layers of documentation generated by successive collection stewards over time and with multiple access controls. References recent research “exploring the perspectives of cultural heritage organisations on Collections as Data frameworks internationally.” Hymbel, Nyhan & Pearlman, Perspectives from Germany and Australia, Collections as Data Infrastructure https://doi.org/10.4324/9781003666530-6.	Ingrid Mason¹ 1. National Museum of Australia
10:45 am - Morning tea (30min)

11:15 am - Keynote (60min)
The Microdata Dividend Abstract At a time when survey response rates are falling globally, the Australian Bureau of Statistics has invested heavily in confidentialised microdata for research. Operated under strict privacy protections and in secure data environments, the Person-Level Integrated Data Asset (PLIDA) and the Business Longitudinal Analytical Data Environment (BLADE) sit alongside other assets such as the Australian Census Longitudinal Dataset, the Longitudinally Linked Employer-Employee Database, the Australian Census and Migrants Integrated Dataset and the Vocational Education and Training National Data Asset. These datasets allow researchers to follow people and firms over time and study policy effects at scale. They enable randomised trials to track outcomes through administrative records and nonprofits to evaluate client outcomes. Already, linked microdata has informed government policies, including merger reform and the decision to ban non-compete clauses for low- and middle-income workers. This talk looks at what Australia’s expanding microdata infrastructure offers researchers, how these assets are being used by international experts, the role of the APS Data Profession, and where the next gains might come from.	Andrew Leigh MP
12:15 pm - Lunch (60min)

1:45 pm - Invited talks: Industry (60min)
Predicting Custodial Length of Stay for Women in NSW Abstract Women in custody experience distinct pathways, needs and offence patterns, and Aboriginal and Torres Strait Islander women are overrepresented in custody. This study investigates whether custodial length of stay can be predicted for women entering custody prior to sentencing using offence, legal history and demographic information. The project is a collaboration between the University of Sydney and Corrective Services NSW, in partnership with MobReady, an Aboriginal-owned organisation supporting women around release from custody. Using Corrective Services NSW’s Offender Integrated Management System and the NSW Bureau of Crime Statistics and Research Re-offending Database, we compare a variety of regression and machine-learning methods to predict length of stay in custody using information available at reception, when approximately 88% of women are unsentenced. Models are evaluated within a consistent predictive framework, and a simulation study is used to validate feature selection methods. The resulting prediction framework could support evidence-based case management, service planning, and pre- and post-release support for women in custody in NSW.	Lily Duck¹ Associate Professor Garth Tarr¹ Dr Evarn Ooi² Lisa Dixon³ 1. School of Mathematics and Statistics, University of Sydney 2. Operational Systems and Integrity, Office of the Commissioner, Corrective Services NSW 3. Aboriginal Strategy and Policy Unit, Corrective Services NSW
A Bayesian Ensemble Projection of Climate Change and Technological Impacts on Future Crop Yields Abstract This work introduces a Bayesian hierarchical modeling framework within a fully probabilistic setting for crop yield estimation, model selection, and uncertainty forecasting under multiple future greenhouse gas emission scenarios. By informing on regional agricultural impacts, this approach addresses broader risks to global food security. Extending an established multivariate econometric crop-yield model to incorporate country-specific error variances, the framework systematically relaxes restrictive homogeneity assumptions and enables transparent decomposition of predictive uncertainty into contributions from climate models, emission scenarios, and crop model parameters. In both in-sample and out-of-sample analyses focused on global wheat production, the results demonstrate significant improvements in calibration and probabilistic accuracy of yield projections. These advances provide policymakers and stakeholders with detailed, risk-sensitive information to support the development of more resilient and adaptive agricultural and climate strategies in response to escalating climate-related risks.	Dan Li¹ Vassili Kitsios² David Newth² Terence O’Kane² 1. Queensland University of Technology 2. CSIRO
DROVER: a beef-sector CGE for supply-shock decisions, run from an ordinary chatbot Abstract Computable general equilibrium models trace shocks through an economy. We built one for Australia’s beef sector, called DROVER. It models herds, feed, processing, and trade, for livestock stakeholders facing a crisis. We work through two decisions. First, an animal phase down against a fuel or food supply interruption known in advance. When feed cannot reach animals, some will be lost, and the model helps size and schedule the phase down to preserve breeding stock and downstream processing capacity. The second aims to avoid the phase down entirely. Here the model plans grain importation, sizing port and storage infrastructure against the cost and lead time of building it. The second thread of the talk is about who can drive the model. This kind of analysis has sat behind a programming barrier that ties the decision-maker to a human modeller. With the Model Context Protocol running in code mode, a user poses scenarios in plain language, revised as questions sharpen, resulting in a policy brief. These tools already sit on every executive’s desktop. In a live session we work from question to brief, then discuss the impact of decision owners manipulating the model themselves.	Alex Cooper Matt van der Saag
Before the Failure: Predictive Fault Detection in Multistage Printhead Manufacturing Abstract Multistage manufacturing processes (MMPs) such as thermal inkjet printhead production accumulate rich measurement data across successive assembly and test operations, yet most quality control remains reactive, identifying defective units only at the final tests. Classical statistical process control (SPC) monitors whether a process remains in control, but does not predict whether an individual unit will fail. This work develops methods for predictive fault detection at the level of the individual unit. A central challenge is selective labelling: units removed at early stages are not observed at later ones, so the available data reflect only those units that survive, biasing conventional predictive models. We investigate approaches based on stream-of-variation (SoV) modelling, conformal prediction and machine learning techniques to detect incipient faults while providing calibrated, unit-level measure of uncertainty. Using data from an industrial manufacturing execution system, we evaluate these methods against standard baselines and discuss their implications for earlier, more reliable fault detection in high-volume manufacturing.	Aaron Gibson¹ Prof Spiridon Penev¹ A/Prof Pierre Lafaye de Micheaux¹ Prof Scott Sisson¹ A/Prof Samuel Mallinson¹ 1. UNSW
2:45 pm - Afternoon tea (30min)

3:15 pm - Invited talks: Environment (60min)
Bayesian Adaptive Sampling Design and Decision Support Tools for Ecologists Abstract Ecological sampling and research prioritisation strategies tend to be guided by local expertise and individual research programs, rather than optimised against explicit, quantitative objectives. A broad family of Bayesian adaptive design methods has the potential to add value across ecological sampling contexts, but remains underutilised. Paired with accessible decision support tools, these methods offer evidence-based approaches to recurring allocation problems in ecology: deciding where and what to sample, which specimens to prioritise for digitisation from collection backlogs, or where to triage conservation information-gathering given national biodiversity priorities. We map the landscape of Bayesian adaptive design methods, distinguishing approaches by goal and utility function - including Value of Information methods, information-theoretic approaches targeting model improvement, and sequential allocation for yield maximisation. Worked examples illustrate practical applications, including expected information gain for wombat sampling design, R√©nyi entropy search for African lovegrass distribution modelling, and species discovery in Australian weevils using bandit algorithms.	Owen Forbes¹ Andrew G. Young¹ Peter H. Thrall¹ Cheng Soon Ong² 1. National Research Collections Australia, CSIRO 2. Technology, CSIRO
Spatial Conservation Prioritisation for Matters of National Environmental Significance Abstract As an authoritative source for national information about Australia’s environment, Environment Information Australia (EIA) publishes geospatial data on where nationally threatened species habitat is likely to occur and how many unique species are predicted to occur in a specific area. This information is used to guide planning and prioritisation of government investment in protection, conservation, and restoration of threatened species and their habitat. Spatial conservation prioritisation (SCPr) integrates the spatial distribution of species with other biodiversity data to produce national-scale prioritisation products. EIA is applying SCPr across multiple use cases including the recently published Relative habitat importance for threatened and migratory species dataset to identify high-priority areas across Australia for nationally threatened biodiversity, offering a consistent and reproducible evidence base to support environmental decision-making. These analyses complement existing regulatory frameworks by providing spatial tools that can inform strategic planning and investment prioritisation in an evolving policy landscape to protect Australia’s most threatened biodiversity.	Dr Kevin D Newman¹ Randall Barry¹ Dr Sarah E Stock¹ Jennifer Dubois¹ Cassandra Malley¹ Dr Dan Rogers¹ Belinda Allison¹ 1. Department of Climate Change, Energy, the Environment and Water
From satellite data to preventative decisions: operationalising remote sensing to reduce sediment pollution risk from urban development Abstract Construction-related earthworks are a persistent source of sediment pollution in urban growth areas, particularly when site preparation and bulk earthworks expose large areas of bare soil on site without appropriate preventative control measures. Regulatory responses are often triggered after environmental harm has occurred. Prevention capabilities are limited by currently available information. This presentation describes how EPA Victoria and University of Melbourne combined satellite imagery, geospatial analytics and environmental science into an operational decision-support capability for proactive, preventative risk-based regulation. The project developed and evaluated a passive, large-scale monitoring system for detecting land disturbance and earthworks at greenfield development sites. The system uses freely available Sentinel-2 imagery and automated analysis of spectral indices to detect bulk earthworks and site preparation activities. The system provides a practical pathway for EPA Victoria decision makers to prioritise inspections, engage duty holders, share intelligence with local government and other partners, and improve erosion and sediment control outcomes.	Adam Gruer¹ Dr Elisa Boera¹ Dr Timothy L Coggan¹ Dr Kathryn Russell² 1. EPA Victoria 2. University Of Melbourne
A Bayesian Multi-Stage Ensemble for Detecting Bats and Raptors from Radar Data Abstract Wind turbine developments in remote arid environments require reliable monitoring of aerial fauna to manage collision risk. Robin Radar systems deployed in the Pilbara region track thousands of airborne objects, but their onboard classifier distinguishes targets mainly by body size, giving limited species-level information. Estimating species-specific collision probability, essential for impact assessment and compliance, depends on resolving this identification gap. This study presents a multi-paradigm ML framework converting radar tracks into species-level classifications. Preprocessing removes clutter and non-biological detections, corrects airspeeds using wind profiles, and adjusts for detection bias. A Bayesian Ensemble integrates unsupervised clustering of body signatures, semi-supervised self-training on flight kinematics, and supervised discriminant analysis of track context, fusing posteriors via Bayesian log-odds with retained uncertainty. The framework identifies micro-bats, ghost bats, and wedge-tailed eagles from nocturnal and diurnal tracks, flagging conflicting detections for review and enabling species-level collision-risk assessment in arid landscapes.	Soudabeh Shemehsavar¹ Patricia A. Fleming² Jill Shephard³ Abderraouf Amrani⁴ Graeme Hocking¹ 1. School of Mathematics, Statistics, Chemistry and Physics, Murdoch University, Perth, WA. 2. School of Environmental and Conservation Sciences, Murdoch University, Perth, WA. 3. Harry Butler Institute, Murdoch University, Perth, WA. 4. School of Information Technology, Murdoch University, Perth, WA.
4:15 pm - Conference close (15min)