A Practical Guide to Selecting Context-Appropriate Fairness Metrics in Machine Learning
Caleb Barr, University Of Canterbury

Recent regulatory proposals for artificial intelligence emphasise fairness requirements for machine learning models. However, precisely defining an appropriate measure of fairness is challenging due to philosophical, cultural and political contexts. Biases can infiltrate machine learning models in complex ways depending on the model’s context, rendering a single common metric of fairness insufficient. This ambiguity highlights the need for criteria to guide the selection of context-aware fairness measures—an issue of increasing importance given the proliferation of ever tighter regulatory requirements of fairness. To address this, we developed a flowchart to guide the selection of contextually appropriate fairness measures. Twelve criteria were used to formulate the flowchart, including the consideration of model assessment criteria, model selection criteria, and data biases. Formulation of this flowchart has significant implications for predictive machine learning models, providing a guide to appropriately quantify fairness in these models.

Pretrained Workflow Transformer: Generating Functional Workflows from Technical Descriptions
Ardi Oeij, Lincoln University

A software functional specification typically includes a technical description of a function without a workflow visualization. Manually creating a workflow of a function from given technical descriptions express in natural language is challenging and time-consuming. One potential solution is to automate the workflow generation. This automation can help System Analyst and Technical Writer to speed up delivery time in creating Software Specification and Technical Documentation.

Earlier studies explored automation using conventional approaches struggled with long sequences data and ambiguity. The introduction of the Transformer model brought a breakthrough in processing long-sequences sequential data and ambiguity. The gap in research lies in the absence of automatic generation of the Workflow for a function from a given technical description by training the Transformer model from scratch. Therefore, this thesis intent to develop a small pretrained transformer model to generate a workflow of a software function

TAIAO Platform: Flood Prediction and Critter monitoring
Nick Lim, AI Institute, University of Waikato

The TAIAO Platform applies artificial intelligence to diverse environmental challenges in New Zealand. This poster highlights two projects: (1) Flood Prediction, which integrates river sensor data, rainfall observations, and machine learning to deliver high-resolution, real-time forecasts that support community resilience against extreme weather; and (2) Maungatautari Species Monitoring, which uses trail cameras and automated image recognition to detect and classify native and invasive species, enhancing biodiversity conservation. Together, these projects illustrate TAIAO’s flexible, data-driven approach and the cross-cutting lessons learned in scaling AI systems, managing heterogeneous datasets, and delivering actionable insights for climate adaptation and ecological stewardship.

Standardizing Cluster Evolution Analysis in Data Streams: CapyMOA as a Platform for Scalable Evaluation
Guilherme Weigert Cassales, AI Institute, University of Waikato

This research focuses on unsupervised clustering in data streams, with an emphasis on understanding and tracking cluster evolution under non-stationary conditions. In streaming environments, clusters may emerge, fade, or shift over time, requiring specialized tools to evaluate and compare algorithmic behavior effectively. To address this, CapyMOA is being developed as a modular and extensible platform that standardizes key aspects of clustering evaluation, including procedures for detecting transitions, benchmarking performance, and ensuring replicability across experiments. By providing a unified API and evaluation framework, CapyMOA aims to facilitate the systematic study of clustering dynamics in high-velocity data, supporting applications in domains such as environmental monitoring, industrial telemetry, and smart infrastructure. This work aims to promote reproducible research and foster broader adoption of stream-based clustering methodologies.

Multi-Kernel CNN Ensemble for Predicting Wildfire Spread
Nikeeta Kumari, Unitec Institute of Technology

Wildfires are now a very serious problem because it is expanding very quickly and cause a lot of damage. They are damaging to the environment, destroy property, put communities at risk, and create serious public health risks by increasing air pollution. Being able to predict how wildfires will spread one day in advance is very important for emergency planning and disaster response. However, this is a difficult task because fire behaviour depends on many different factors, such as weather, vegetation, terrain, and past fire activity.

This project focuses on using a Multi-Kernel Convolutional Neural Network (MKCNN) to predict the next-day wildfire spread. The model will use satellite-based environmental image data such as temperature, humidity, wind direction, vegetation, and elevation. The goal is to predict the fire extent for the following day.

The dataset “Next Day Wildfire Spread”, which I will use, includes daily fire masks along with weather and environmental features. The data is in a structured grid, which makes it suitable for deep learning models like CNNs. Since this dataset is standardised and widely used, I can compare results with other studies. At the same time, there are challenges. Most pixels show no fire, so the data is imbalanced and harder to train on. In addition, since this dataset is still relatively new and not much research has been done with it, there may be unexpected challenges during the project, which I am prepared to address.

Rather than training a single model with all features at once, the proposed approach trains several MKCNN models using pairs of inputs: one environmental feature combined with the previous day’s fire mask. An extension of this approach could be the training of models with a fire mask and two of the top-ranked features. The final result will be calculated by ensembling these trained models to improve accuracy and stability.

The performance of this method will be compared with a baseline eleven-feature MKCNN model trained on all features, as well as other models such as Convolutional Autoencoders and U-Nets, which already showed their high performance in the prediction of wildfires. The project also finds which environmental factors have a high impact on the prediction of wildfire spread. It will provide information that could support better decision-making in wildfire management.

On the Shoulders of the Internet: Unveiling the Mystique of AI and Power
Lisen He, Victoria University of Wellington

AI is never neutral; it is inherently a power phenomenon. Understanding AI and its societal impacts requires a comprehensive analysis of its relationship with power. For this emerging field of research, despite diverse investigations into the nexus of AI and power, a systematic and unified description of their relationship is still lacking in the current literature at its early stage of development. Without such conceptual clarity, it is difficult to effectively confront the rise of AI hegemony in the digital era. To address this gap, this literature-based research thesis aims to develop a unifying conceptual framework that captures the overall relationship between AI and power. This framework serves as an integrative synthesis, encompassing relevant studies across various domains and offering a coherent structure to systematically map this diverse body of work. It allows scholars to situate their individual research within this large picture. On this basis, this thesis further aims to redefine the concept of power through a critical review. This thesis consists of two major components: (1) the development of the conceptual framework, and (2) a novel classification approach of the concept of power in the context of AI. Unlike traditional approaches that rely on aggregating prior studies, this thesis draws inspiration from the evolution of Internet Studies to inform AI Studies. It adapts and updates theoretical insights from Internet research to the AI context, while also proposing new theoretical implications specific to AI. These together constitute an eventual finding – the conceptual framework of the relationship between AI and power. This thesis argues that a true understanding of the relationship between AI and power requires basing on this conceptual framework (by recognizing both the formation and practice of AI at the same time).

Theoretically, this thesis aims to provide a comprehensive understanding of the relationship between AI and power, and to offer a broad conceptual framework through which scholars can resituate their work. While the thesis offers limited empirical insights, its primary contribution lies in advancing conceptual and theoretical understandings within the field of Science, Technology, and Society Studies.

Cross-Cultural Comparison on the Ethical Adoption of Artificial Intelligence between Malaysia and New Zealand
Asma Mat Aripin, Massey University

The rapid adoption of Artificial Intelligence (AI) presents complex ethical challenges that vary across cultural and regulatory contexts. This research explores how Malaysia and New Zealand, two nations with distinct cultural values and governance frameworks, approach the ethical adoption of AI in healthcare, business, and government sectors. Drawing on 42 semi-structured interviews (20 from Malaysia, 22 from New Zealand) and thematic analysis using NVivo, the study identifies key similarities and divergences in themes such as transparency, privacy, accountability, and trust. Findings highlight that while both countries recognize the need for robust governance and ethical safeguards, New Zealand emphasizes rights-based frameworks and regulatory compliance, whereas Malaysia prioritizes collective values and socio-cultural acceptance in AI deployment. This comparative perspective offers critical insights for policymakers, practitioners, and academics by demonstrating the importance of contextualized AI governance. The research contributes to cross-cultural scholarship on digital ethics and provides actionable recommendations to support equitable and responsible AI adoption in diverse societies.

Stability driven reinforcement learning for robotic construction
Marcel Garrobe, AI Institute, University Of Waikato

This project explores the use of reinforcement learning for autonomous robotic construction of stable structures without scaffolding. The focus lies on designing reward functions that effectively guide a robotic agent to connect two fixed points by placing blocks. Two metrics were developed to evaluate structural stability: one based on the maximum vertical load each block can bear, and another based on the critical tilting angle before collapse. These metrics were used to shape more informative reward signals compared to traditional binary schemes. Results show that this approach improves learning efficiency and leads to the construction of more robust and reliable designs.

Quantum Re-Uploading for Streaming Time-Series: Fourier Analysis and Climate-Risk Use-Cases
Léa Cassé, AI Institute, University of Waikato

Quantum Re-Uploading Units (QRUs) are shallow, hardware-efficient quantum models that repeatedly encode inputs through a single qubit, trading circuit depth and entanglement for richer frequency representations. This work examines their expressivity and trainability through Fourier spectral analysis and an absorption-witness metric, showing how re-uploading depth L shapes accessible frequencies while mitigating gradient instabilities typical of deeper ansätze. Performance is compared with parameter-matched classical baselines (LSTMs and MLPs) on the Mackey-Glass benchmark and TAIAO river-level datasets, highlighting stable training and enhanced spectral diversity. An applied prototype coupling a QRU forecaster with a QAOA-CVaR allocator demonstrates a practical use case for parametric micro-insurance, using TAIAO environmental and MetService data. Presented within the Global Industry Challenge (World Bank track) and Quantum World Congress 2025, this study provides a reproducible framework connecting spectral diagnostics with hybrid quantum-classical models such as QLSTM or quantum-enhanced reinforcement learners.

Adaptive Isolation Forest
Justin Liu, AI Institute, University Of Waikato

Anomaly detection in real-world data streams often struggles with concept drift, where evolving data distributions challenge algorithms to maintain a balance between accuracy, speed, stability, and plasticity. We present Adaptive Isolation Forest (AIF), a novel anomaly detection algorithm designed to effectively adapt to such changes in a resource-efficient and balanced manner. AIF's novelty can be found in the combination of a smart model update mechanism together with a newly developed MinTreeMaxMass (MTMM) criterion, which scores individual trees for replacement. Extensive evaluations on various benchmark datasets demonstrate that AIF significantly outperforms existing state-of-the-art streaming anomaly detection algorithms in terms of detection accuracy. Moreover, AIF achieves linear time and space complexities, providing a robust solution that maintains high accuracy and efficiency, balancing stability and plasticity in dynamic data streams.

A Large Language Model for Te Reo Māori
Luca Blaauw Fossen, AI Institute, University Of Waikato

Large language models (LLMs) have achieved remarkable performance across natural language processing tasks, yet frontier models perform poorly for te reo Māori and other low-resource languages. The strongest existing models are also proprietary, inaccessible to researchers, and trained on data of unclear provenance. Moreover, these globally optimised models often fail to reflect Māori and Pasifika linguistic and cultural contexts.

We aim to develop the first sovereign Māori large language model. This initiative involves curating high-quality Māori text datasets, constructing culturally relevant benchmarks, and pretraining and instruction-tuning a foundation LLM. This will be done under Māori expert oversight and community participation from iwi and Māori speakers.

We will perform continual pretraining (CPT) and instruction fine-tuning (IFT) of frontier models such as Llama and Mistral. Novel methods of improving data and compute efficiency will be explored, including data augmentation, tokenizer adaptation and parameter-efficient fine-tuning (PEFT). Evaluation will combine automatic metrics with leaderboard-based and expert assessments to capture performance and cultural competency.

This will be the first sovereign Māori foundation model, enabling general-purpose and culturally aligned NLP applications. Beyond Māori, this work will support the development of LLMs for other low-resource languages.

AutoSAD: An Adaptive Framework for Streaming Anomaly Detection
Nilesh Verma, AI Institute, University of Waikato

Real-time anomaly detection in data streams requires continuous adaptation to evolving patterns and concept drift, yet existing methods rely on static algorithm selection and fixed hyperparameters that become suboptimal as data characteristics change. We introduce AutoSAD, the first fully autonomous framework that solves unsupervised streaming anomaly detection through intelligent model selection. Our approach maintains an ensemble of diverse detectors and employs multi-armed bandit optimization with normalized anomaly scores as reward signals, coupled with evolutionary hyperparameter mutation guided by performance feedback. Comprehensive evaluation on diverse datasets demonstrates that AutoSAD achieves superior performance, outperforming state-of-the-art streaming detectors and showing statistically significant improvements across varying data stream characteristics.

Real-Time NZ sign language translator using Machine Learning
Mathes Kankanamge Chami Jayasekera, University of Waikato

New Zealand Sign Language (NZSL) presents distinct challenges for automated translation because it requires the simultaneous analysis of both hands' gestures to decode meaning. The NZSL gesture translator can improve the living standards of impaired communities in New Zealand. This study aims to assess the success of translating NZSL into text by training an image recognition system using a Convolution Neural Network (CNN).

NZSL alphabet images were captured using a 720p camera and were processed using a Convolutional Neural Network (CNN) with transfer learning on an NVIDIA Jetson Nano platform, which was enabled for on-device pre-processing and real-time inference. Video recordings of gestures were segmented into single frames, producing 87,000 labelled images across 29 gesture classes. The data were pre-processed using normalization and augmentation (rotation, scaling, flipping). The training algorithm was developed in Python using the TensorFlow framework, employing an InceptionV3 model based on CNN with transfer learning. The model was optimized using the Adam optimizer with a batch size of 32 and a learning rate of 0.001.
Experimental evaluation was conducted, and 99.1% test accuracy was achieved, with classification scores across all gesture classes ranging from 76% to 98% and F1-scores between 88–100%. The Cohen’s kappa coefficient was calculated as 0.957, indicating a very high level of agreement between predicted and true labels. Performance validation using confusion matrices, training/validation accuracy, and loss curves was performed, and it was observed that the model converged stably, with minimal overfitting and consistently low cross-entropy values. On the Jetson Nano, real-time inference was achieved at approximately 30-40 frames per second (fps), demonstrating its practical deployment potential.

The results demonstrate that individual NZSL alphabet signs can be classified with high accuracy, a critical step towards the long-term goal of translating continuous sign language into text. Results indicate that NZSL sign language can be accurately translated to digital text using computer vision technology. A primary limitation is that the system’s performance is susceptible to variations in lighting conditions and background context. Nevertheless, these findings highlight the feasibility of applying computer vision models to NZSL, offering promising applications in education, healthcare, and public services to enhance communication with the hearing-impaired community.

These results demonstrate that NZSL gestures are technically more complex due to their dual-handed structure and spatiotemporal interdependence (unlike single-handed sign languages), can be accurately translated with computer vision models, as demonstrated by these results.

A new UNet transfer function to significantly improve consistency of image Segmentation accuracy across breast tumour types and patients
Reza Tayeh, Lincoln University

Breast cancer is the most common cancer among women (WHO, 2025). In New Zealand alone, approximately 3,500 new cases are reported each year, with around 650 annual deaths in 2020, according to the latest report from the Ministry of Health in 2023 (Health New Zealand, 2025). The survival rate exceeds 90% when the disease is diagnosed at an early stage. Therefore, accurate and reliable diagnosis, as well as precise segmentation of the tumour location, is crucial for improving survival rate. Artificial intelligence (AI) can play a vital role in this regard.

Many studies have been conducted on segmenting breast tumours from medical images such as ultrasound scans using AI to assist doctors in achieving fast and reliable segmentation. Recently, a new convolutional neural network (CNN) model, called Attention U-Net, has been applied for breast tumour segmentation. Although some studies reported high accuracy of 98% (Mridha et al., 2023) or Dice scores of 97% for tumour detection (Prasetyo, 2024) on ultrasound image segmentation during training, these results were not consistent across all tumour types. Furthermore, these studies did not provide sufficient data regarding model testing.

To evaluate the model more thoroughly, the Attention U-Net was simulated again on the same ultrasound images using the same configuration as above latest models which is done by (Prasetyo, 2024). The results showed that the model provided high accuracy for some tumours, but for others, accuracy was not high enough. Although the maximum Dice score reached approximately 98%, in 91 out of 113 patient test images, the Dice score was below 90%. Further, among the 113 test patients, the Dice score was below 80% in 43 patients, below 70% in 43 patients, and below 50% in 21 patients. These findings indicate that while the model can achieve high accuracy scores in some cases, its performance is not reliably consistent across different tumour types or patients.

Consistency is crucial in medical imaging because it directly affects patients’ lives. In an effort towards achieving this goal, in this study, the parameters of the Attention U-Net were adjusted to improve model accuracy across the board. A breakthrough success however was finally achieved by designing a new activation function called Parametric Tanh. With this new activation function, the highest Dice score increased to approximately 99%. On the same test set of 113 patients, the Dice score was above 90% in 66 patients and above 80% in 81 patients, compared to 22 and 70, respectively, in previous studies. With our model, only 21 patients had Dice scores below 70%, and just 11 patients out of 113 had segmentation accuracy below 50%, compared to 43 and 21 patients, respectively, in previous studies. These results represent increases of 200% for over 90% accuracy, 16% for over 80% accuracy, 104% improvement for below 70% accuracy and 91% for below 50% accuracy- an overall accuracy increase for 87 out of 113 patients (77%) in the test set. Thus, with the proposed activation function, the model can achieve significantly more reliable tumour segmentation across different tumour types and patients.

An AI driven approach for flood mapping and impact forecasting at household level in New Zealand
Malintha Mahinda Kumarage, Massey University

Flooding is New Zealand’s most frequent natural hazard, regularly causing socio-economic impacts and highlighting an urgent need for more effective flood warnings. Impact-Based Forecasts and Warning Services (IBFWS), endorsed by the World Meteorological Organisation (WMO), aim to provide more actionable information to decision makers and at-risk communities by focusing not only on the expected hazard but also on its potential impacts. However, the practical implementation of IBFWS for floods remains challenging, particularly in New Zealand. A key barrier is the need for rapid hazard and impact modelling within the short forecast lead time (often only a few hours) of weather events, which is computationally intensive using c onventional modelling approaches.

This research addresses this modelling challenge by developing ‘HydroNetNZ’, an experimental household-level IBFW system for floods in New Zealand. The system employs a deep learning (DL)-based emulator modelling approach to forecast high-resolution flood inundation maps, which are then translated to household-level impact forecasts using a novel machine-learning-based method. The ‘HydroNetNZ’ mobile app, currently under development with ongoing engagement with key stakeholders including emergency managers and hydrologists, will facilitate a user-centric implementation of this experimental IBFW system across multiple case study areas, ultimately improving the resilience of flood-prone communities in New Zealand.

Beyond the Hype: Alignment on AI Risks Between Developers and Managers
Mike Watts, Media Design School at Strayer

Coding assistants powered by artificial intelligence like GitHub Copilot are quickly transforming software building workflows. The popular narrative around their adoption is that of conflict: productivity boosts are invited by developers, but managers reject it on grounds of security and compliance risk. This narrative has resulted in governance constructs developed on supposed tension instead of empirical observation.

This research put that assumption to the test with empirical data. Based on a questionnaire of 100 software professionals such as developers, senior engineers, and technical managers it surveys judgments about coding assistants’ effect on productivity, code quality, and security.

Results confirm the broadly reported productivity gains: respondents largely concurred that coding is sped up by AI tools and development time shrinks. More importantly, though, the results revealed a surprising consensus: managers and developers alike reported moderate-to-high concern about data security threats. Instead of rival groups, results showed a mutual “trust but verify” attitude that accepts AI’s worth but places strict emphasis on human verification checks.

By debunking the myth of a developer - manager conflict, this study points to a missed opportunity. The current alignment around risk perception forms a basis on which organizations can craft AI adoption plans that focus on cooperation rather than conflict. This evidence-based outlook changes the discussion from managing compromises to building a shared culture of innovative and responsible use of AI.

Federated Learning on Bee Acoustic and Environment Data for Generalisable Environmental Monitoring
Saba Mustafa, Auckland University of Technology

Honeybee colonies are vulnerable to environmental fluctuations, and their acoustic emissions provide valuable bioindicators of stress. Our research addresses a critical challenge in environmental monitoring: predicting wildfire risk and ecological stress using bioacoustics signals from bees. Building on recent systematic literature review findings, highlighting the combined effect of temperature and humidity on wildfire patterns and bee colony health, we designed experiments to quantify these environmental effects using hive audio data.

We extracted Mel-Frequency Cepstral Coefficients (MFCCs) from audio recordings of live bee colonies and systematically analyzed their correlation with in-hive temperature and humidity. The experiments show a robust, statistically significant relationship: Random Forest models predicted humidity (R²=0.98, RMSE=2.18%) and temperature (R²=0.96, RMSE=1.44°C) from MFCC features with high accuracy. Feature importance analyses and curve panels further reveal that specific MFCC components (such as mfcc08_mean and mfcc00_mean) are especially sensitive to these environmental variables, supporting the idea that bee acoustic patterns encode real-time physical conditions within the hive.

Building on these results, we propose a federated learning (FL) framework to overcome significant challenges such as data scarcity and site heterogeneity in developing generalizable models. In this decentralized AI approach, models are trained locally at each hive or monitoring site, and only model parameters are shared and aggregated. This approach keeps data private, uses data from many different sources, and makes the model stronger to find patterns in bee and environmental data from many different places and climates.

This study aims to establish a connection between hive sounds and environmental stress and provide a way to expand bee sound monitoring using federated learning. This approach helps solve the problem of limited data, makes the model more generalizable for pattern detection, and opens new ways to use bee sounds for real-world, privacy-preserving wildfire and environmental monitoring.