Deep neural networks often exhibit degraded performance under domain shifts due to reliance on domain-specific features. Existing domain generalization (DG) methods attempt to mitigate this during training but lack mechanisms to adaptively correct domain-specific reliance once it emerges. We propose Identify and Unlearn (IU), a model-agnostic module that continually mitigates such reliance post-epoch. We introduce an unlearning score to identify training samples that disproportionately increase model complexity while contributing little to generalization, and an Inter-Domain Variance (IDV) metric to reliably identify domain-specific channels. To suppress the adverse influence of identified samples, IU employs a Domain-Specific Gradient-Ascent (DSGA) procedure that selectively removes domain-specific features while preserving domain-invariant features. Extensive experiments across seven benchmarks and fifteen DG baselines show that IU consistently improves out-of-distribution generalization, achieving average accuracy gains of up to 3.0\%.

Introduction:Timely identification of dementia remains a major clinical challenge globally, with many cases being unrecognized. This study evaluated whether using machine learning models with routinely collected health data can support dementia case finding. Method:De-identified datasets were used to create a nested case-control of 8195 individuals with dementia and 8195 matched controls. Four models incorporating both cross-sectional and longitudinal features were developed and tested. Results:The best-performing model achieved an area under the curve of 0.86 (95% confidence interval [CI]: 0.84-0.87), with sensitivity of 73.3% (95% CI: 72.3-74.2) and specificity of 87.5% (95% CI: 86.7-88.2). Key predictors included time-stamped International Classification of Diseases 10th Revision diagnostic codes, health-care use, referral to aged residential care, and hospital delirium assessments. Discussion:A unique feature was the inclusion of "timestamp data" that allowed us to assess the longitudinal changes which may have improved performance of the model. These findings demonstrate the potential for using machine learning with routine health data to enhance early dementia detection. Highlights:Machine learning models using routine health data in a real-world setting in New Zealand had good accuracy for identification of dementia (area under the curve 0.86, sensitivity 73.3%, and specificity 87.5%).Longitudinal sequential "timestamp" data is a unique feature that improved the performance of machine learning models for dementia case finding.Key predictors included International Classification of Diseases 10th Revision codes, health-care use, referrals to aged residential care, and positive delirium scores.

Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, highlighting the critical need for accurate and interpretable early risk prediction. While conventional tools such as the Framingham Risk Score offer high transparency, they rely on rigid assumptions and fixed variable sets, limiting their adaptability to patient-specific variations. Recent machine learning approaches improve predictive performance but often lack interpretability, hindering clinical acceptance. To address these limitations, we propose CluRF, a hybrid framework that combines unsupervised clustering with Random Forest classifiers and integrates SHAP-based explanations. CluRF partitions the population into latent subgroups based on feature similarity, enabling personalized modeling within each cluster and providing global and individualized interpretability. We evaluate our method on four real-world heart disease datasets and demonstrate that CluRF achieves competitive or superior performance compared to state-of-the-art models in terms of F1-score, accuracy, and AUC. Importantly, it maintains a lightweight and explainable structure suitable for clinical deployment.

Dementia is a major global public health challenge, affecting approximately 55 million people worldwide, with an estimated 10 million new cases annually (WHO). In New Zealand, prevalence estimates based on national datasets suggest that dementia affects 3.8%–4.0% of individuals aged 60 and older. When accounting for undiagnosed cases using a capture-recapture method, this estimate increases to 9.2% (95% CI: 8.9%–9.6%), with disproportionately higher rates among Māori and Pacific populations. However, nearly 50% of dementia cases remain undiagnosed, limiting timely interventions and increasing healthcare costs. There is an urgent need for scalable, data-driven solutions to improve dementia identification. This study aimed to develop accurate machine learning models for dementia identification using routinely collected health data. Routinely collected health data from the Te Whatu Ora Counties Manukau population (aged 65+) were analysed, incorporating sociodemographic and clinical variables (both longitudinal and cross-sectional). Dementia status was determined based on pharmacy records (antidementia drug prescriptions), interRAI assessments (dementia-related evaluations), and hospitalization records (ICD-10 codes related to dementia). A nested one-to-one case-control design was implemented, retaining pre-diagnostic information across six-time windows before diagnosis (0 days, 6 months, 1 year, 3 years, 5 years, and 8 years). Deep learning models were trained using a training/validation/testing framework. The models achieved an accuracy of 80.47% (95% CI: 80.23–80.72) when using data immediately before diagnosis with a sensitivity of 71.77% (95% CI: 71.73–71.81) and specificity of 87.27% (95% CI: 87.24–87.30). The highest performance was achieved when all available features were included. Key predictive features included Aged Residential Care (ARC)-related factors, comorbidities (ICD-10 codes), and ethnicity. Among pharmacy-related variables, analgesics, diuretics, antithrombotics, anti-epileptic drugs, and diabetes treatments were highly relevant. The number of ophthalmology and cardiology appointments, along with delirium-related features, also contributed significantly. Machine learning applied to routinely collected health data offers a powerful, scalable approach to dementia case-finding, improving early identification and facilitating timely interventions. These models have the potential to enhance clinical decision-making, optimize healthcare resources, and ultimately improve outcomes for individuals living with undiagnosed dementia.

A design model is the abstract representation of an actual process or software product. Although some software faults can be found by diagnosing design models before implementation, repairing the design models is time-consuming to software developers. To achieve faster software development, this paper introduces an automated approach to generally repair design models diagnosed by model checking. Model checkers are used to detect faults such as unreachable goals and violated properties in design models. Such faults are eliminated in parallel by insertion, modification and deletion operators found by constraint solving and predictive models. The outcomes of model repair are evaluated using the ISO/IEC 25010 software quality metrics. Experimental results have demonstrated that the proposed approach can eliminate unreachable goals and invariant violations in various design models while preserving their model quality. The effectiveness and performance of such design model repair processes depend mainly on the complexity of design model, the efficiency of constraint solver and the accuracy of predictive model. This study indicates that model-driven software development can be more efficient by automating model diagnosis, fault elimination and quality evaluation.

In recent years, the integration of machine learning techniques into chemical reaction product prediction has opened new avenues for understanding and predicting the behaviour of chemical substances. The necessity for such predictive methods stems from the growing regulatory and social awareness of the environmental consequences associated with the persistence and accumulation of chemical residues. Traditional biodegradation prediction methods rely on expert knowledge to perform predictions. However, creating this expert knowledge is becoming increasingly prohibitive due to the complexity and diversity of newer datasets, leaving existing methods unable to perform predictions on these datasets. We formulate the product prediction problem as a sequence-to-sequence generation task and take inspiration from natural language processing and other reaction prediction tasks. In doing so, we reduce the need for the expensive manual creation of expert-based rules.

Continual Learning (CL) methods enable models to learn new tasks without forgetting previously learned ones. Catastrophic Forgetting (CF) occurs when the parameters of a neural network are updated for a new task, causing the model to lose performance on tasks it has previously learned. To mitigate CF, parameter isolation methods use a “task mask” to allocate a subset of weights to each task; these weights are typically frozen to preserve task performance. However, frozen weights can limit positive backward transfer, which is the beneficial reuse of knowledge from new tasks to improve the accuracy of previously learned tasks. To address this gap, we introduce LEarning AFter learning (LEAF), a novel CL method that enables positive backward transfer by dynamically updating frozen task masks based on gradient updates that signal sufficient backward knowledge transfer. This mechanism allows for selective integration of new knowledge without sacrificing previously acquired knowledge. Our experiments show that LEAF surpasses existing state-of-the-art methods in terms of accuracy while maintaining comparable memory and runtime efficiencies. Moreover, it outperforms other backward transfer techniques in improving the accuracy of a prioritized task. Our code is available at https://github.com/wernse/LEAF.

Traditional deep learning models often struggle in few-shot learning scenarios, where limited labeled data is available. While the Contrastive Language-Image Pre-training (CLIP) model demonstrates impressive zero-shot capabilities, its performance in few-shot scenarios remains limited. Existing methods primarily aim to leverage the limited labeled dataset, but this offers limited potential for improvement. To overcome the limitations of small datasets in few-shot learning, we introduce a novel framework, SSAT-Adapter, that leverages CLIP's language understanding to generate informative auxiliary tasks and improve CLIP's performance and adaptability in few-shot settings. We utilize CLIP's language understanding to create decision-boundary-focused image latents. These latents form auxiliary tasks, including inter-class instances to bridge CLIP's pre-trained knowledge with the provided examples, and intra-class instances to subtly expand the representation of target classes. A self-paced training regime, progressing from easier to more complex tasks, further promotes robust learning. Experiments show our framework outperforms the state-of-the-art online few-shot learning method by an average of 2.2% on eleven image classification datasets. Further ablation studies on various tasks demonstrate the effectiveness of our approach to enhance CLIP's adaptability in few-shot image classification.

Federated learning (FL) is a popular approach to facilitate privacy-aware machine learning since it allows multiple clients to collaboratively train a global model without granting others access to their private data. It is, however, known that FL can be vulnerable to membership inference attacks (MIAs), where the training records of the global model can be distinguished from the testing records. Surprisingly, research focusing on the investigation of the source inference problem appears to be lacking. We also observe that identifying a training record's source client can result in privacy breaches extending beyond MIAs. For example, consider an FL application where multiple hospitals jointly train a COVID-19 diagnosis model, membership inference attackers can identify the medical records that have been used for training, and any additional identification of the source hospital can result the patient from the particular hospital more prone to discrimination. Seeking to contribute to the literature gap, we take the first step to investigate source privacy in FL. Specifically, we propose a new inference attack (hereafter referred to as source inference attack -- SIA), designed to facilitate an honest-but-curious server to identify the training record's source client. The proposed SIAs leverage the Bayesian theorem to allow the server to implement the attack in a non-intrusive manner without deviating from the defined FL protocol. We then evaluate SIAs in three different FL frameworks to show that in existing FL frameworks, the clients sharing gradients, model parameters, or predictions on a public dataset will leak such source information to the server. We also conduct extensive experiments on various datasets to investigate the key factors in an SIA. The experimental results validate the efficacy of the proposed SIAs.

Despite the significant advances in supervised person re-identification (ReID) methods, these models exhibit performance degradation in unseen domains. Domain generalization (DG) is applied to alleviate this issue, but most existing DG methods assume consistent class spaces between source and target domains. We propose Adaptive Adversarial Augmentation (AAA), a Heterogeneous Domain Generalization (HDG) approach tailored for single-source cross-dataset ReID. AAA jointly trains a feature extractor alongside a Domain Adversarial Network (DAN) and a Class Adversarial Network (CAN) to enhance the feature extractor's robustness to both domain shifts and class space changes. Additionally, we propose a diversity-based perturbation impact factor, dynamically tuning the perturbation influence aligned with the diversity of learned embeddings, thus providing a flexible augmentation strategy. Experimental results demonstrate that our method surpasses state-of-the-art methods on large-scale cross-dataset ReID benchmarks.

Adversarial defenses protect machine learning models from adversarial attacks, but are often tailored to one type of model or attack. The lack of information on unknown potential attacks makes detecting adversarial examples challenging. Additionally, attackers do not need to follow the rules made by the defender. To address this problem, we take inspiration from the concept of Applicability Domain in cheminformatics. Cheminformatics models struggle to make accurate predictions because only a limited number of compounds are known and available for training. Applicability Domain defines a domain based on the known compounds and rejects any unknown compound that falls outside the domain. Similarly, adversarial examples start as harmless inputs, but can be manipulated to evade reliable classification by moving outside the domain of the classifier. We are the first to identify the similarity between Applicability Domain and adversarial detection. Instead of focusing on unknown attacks, we focus on what is known, the training data. We propose a simple yet robust triple-stage data-driven framework that checks the input globally and locally, and confirms that they are coherent with the model’s output. This framework can be applied to any classification model and is not limited to specific attacks. We demonstrate these three stages work as one unit, effectively detecting various attacks, even for a white-box scenario.

In continual learning, a primary factor of catastrophic forgetting is task-recency bias, which arises when a model is trained on an imbalanced set of new and old task instances. Recent studies have shown the effectiveness of rehearsal-based continual learning methods; however, a major drawback of these methods is the loss of accuracy on older tasks when training is biased towards newer tasks. To bridge this gap, we propose a λ Stability Wrapper ( λ SW), where the learner uses a task-based policy to adjust the probability of when instances are replaced in memory to account for task-recency bias to alleviate catastrophic forgetting. The policy results in an increased number of instances seen from older tasks. By construction, λ SW can be applied with other rehearsal-based continual learning algorithms. We validate the effectiveness of λ SW with three well known baseline methods: Gradient-based Sample Selection, Experience Replay, and Maximally Interfered Retrieval. Our experimental results show significant gains in accuracy on eleven out of twelve of our experiments across four datasets.

The automation of programming, which lies at the intersection of software engineering and artificial intelligence, enables machines to automatically generate programs that satisfy given requirements. In the context of B formal design modeling, one of the challenges is the refactoring of substitutions in design specifications, which often uses state transitions to describe how program or system statuses change during execution. This paper proposes a condition and substitution refactoring algorithm for the B formal specification language. The aim of the work is to automatically derive B operational predicates based on given transitions. The work has been extremely useful to machine-driven formal design model repair as well as automated design specification generation. Given a set of state transitions, common relations of their state variables can be discovered and clustered into a number of classes. These relations can be further used to synthesize substitutions that derive new states from existing states. To restrict application domains of the synthesized substitutions, conditions that guard these substitutions are generated using first-order logic. We have implemented the proposed algorithm as an extension to the ProB model checker. Experiments were conducted based on the B model public dataset. The evaluation results demonstrated that our solution is able to synthesize conditions and substitutions for various sets of state transitions in a wide range of B models.

Time Series Forecasting (TSF) is well established in domains dealing with temporal data to predict future events yielding the basis for strategic decision-making. Previous research indicated that forecasting models are vulnerable to adversarial attacks, that is, maliciously crafted perturbations of the original data with the goal of altering the model’s predictions. However, attackers targeting specific outcomes pose a substantially more severe threat as they could manipulate the model and bend it to their needs. Regardless, there is no systematic approach for targeted adversarial learning in the TSF domain yet. In this paper, we introduce targeted attacks on TSF in a systematic manner. We establish a new experimental design standard regarding attack goals and perturbation control for targeted adversarial learning on TSF. For this purpose, we present a novel indirect sparse black-box evasion attack on TSF, n Vita. Additionally, we adapt the popular white-box attacks Fast Gradient Sign Method (FGSM) and Basic Iterative Method (BIM). Our experiments confirm not only that all three methods are effective but also that current state-of-the-art TSF models are indeed susceptible to attacks. These results motivate future research in this area to achieve higher reliability of forecasting models.

Extensive labeled training data for anomaly detection is enormously expensive and often unavailable in data-sensitive applications due to privacy constraints. We propose TransForest, a transductive forest for anomaly detection, in the semi-supervised setting where few labels are available. Guided by little label information, TransForest pushes classification boundaries toward sensitive areas where abnormal and normal points are located, increasing learning capacity. Empirically, TransForest is competitive with other unsupervised and semi-supervised representative detectors given a small number of labeled points. TransForest also offers a feature importance ranking consistent with the rankings provided by popular supervised forests on low-dimensional data sets. Our code is available at https://github.com/jzha968/transForest .

Introduction Developments in Artificial Intelligence (AI) are adopted widely in healthcare. However, the introduction and use of AI may come with biases and disparities, resulting in concerns about healthcare access and outcomes for underrepresented indigenous populations. In New Zealand, Māori experience significant inequities in health compared to the non-Indigenous population. This research explores equity concepts and fairness measures concerning AI for healthcare in New Zealand. Methods This research considers data and model bias in NZ-based electronic health records (EHRs). Two very distinct NZ datasets are used in this research, one obtained from one hospital and another from multiple GP practices, where clinicians obtain both datasets. To ensure research equality and fair inclusion of Māori, we combine expertise in Artificial Intelligence (AI), New Zealand clinical context, and te ao Māori. The mitigation of inequity needs to be addressed in data collection, model development, and model deployment. In this paper, we analyze data and algorithmic bias concerning data collection and model development, training and testing using health data collected by experts. We use fairness measures such as disparate impact scores, equal opportunities and equalized odds to analyze tabular data. Furthermore, token frequencies, statistical significance testing and fairness measures for word embeddings, such as WEAT and WEFE frameworks, are used to analyze bias in free-form medical text. The AI model predictions are also explained using SHAP and LIME. Results This research analyzed fairness metrics for NZ EHRs while considering data and algorithmic bias. We show evidence of bias due to the changes made in algorithmic design. Furthermore, we observe unintentional bias due to the underlying pre-trained models used to represent text data. This research addresses some vital issues while opening up the need and opportunity for future research. Discussions This research takes early steps toward developing a model of socially responsible and fair AI for New Zealand's population. We provided an overview of reproducible concepts that can be adopted toward any NZ population data. Furthermore, we discuss the gaps and future research avenues that will enable more focused development of fairness measures suitable for the New Zealand population's needs and social structure. One of the primary focuses of this research was ensuring fair inclusions. As such, we combine expertise in AI, clinical knowledge, and the representation of indigenous populations. This inclusion of experts will be vital moving forward, proving a stepping stone toward the integration of AI for better outcomes in healthcare.

In the online data stream environment, a model is typically not effective until a sufficient number of data instances have been seen. One solution to alleviate this issue is using model transfer. A major gap in current research is that they do not address the cost-effectiveness of model transfer in the online context, where processing time is crucial as data instances continuously arrive at high-speed. Model transfer in data streams involves continuously adapting the model to new data in the target stream. Suppose the target stream contains hard-to-learn patterns not covered by the transferred model. In that case, adaptation may incur extra processing time with little accuracy performance gains compared to simply building a new model for the target stream from scratch. Therefore, considering the cost of model transfer and target model construction is essential to balance the accuracy performance gains against computation cost for cost-effective model transfer. To address this gap, we propose a framework called OPERA (Online Transfer using Phantom Tree for Real-Time Adaptation), that orchestrates transfer learning based on the accuracy gain and runtime tradeoffs between transferring and adapting a source model, versus no transfer and constructing a new model for the target stream. We conduct extensive empirical studies to show that our framework can balance accuracy performance gains and runtime for cost-effective transfer learning in the data stream environment.

Pollution from wood burners has profound health implications for the general population. Typically, monitoring the level of airborne particulate matter, PM 2.5 , in these areas often requires making inferences about missing or corrupted readings. Air Quality inference in these cases often poses critical challenges. The factors can evolve over time, changing the distribution of data. Such changes in the distribution of data are known as concept drift. Moreover, air pollution inference for a location typically would require historical data to be collected for the location. We investigate five air quality studies in New Zealand rural towns. We explore two different research problems: (1) an adaptive recurrent drift algorithm to model recurrence patterns in PM 2.5 levels for a town with the ability to recover after accuracy deterioration after a concept drift using an adaptive recurrent drift algorithm, and (2) transfer learning for the data stream whereby we reuse a pre-trained air pollution inference model from a town as the starting point for an air pollution inference model on another town. We further investigate the relationship between the changes we detected and changes within the prediction horizon. We showed that the average accuracy of the air quality inference for the five towns is between 70% and 94% using the recurrent drift algorithm. We also show that transfer learning was advantageous between two of the five towns.

Most machine learning algorithms rely on a set of hyperparameters, which are tuned either by domain experts or using automated methods. In the context of data streams, the complexity of this task is compounded by the need for tuning every time a concept drift occurs. The problem of automated tuning may lead to classification performance improvements. Current methods optimise the hyperparameters from scratch every time concept drift is detected. Additionally, meta-learning techniques developed for static machine learning tasks allow pre-built knowledge to adapt faster to new tasks. We propose a meta-learning approach for automated tuning in evolving data streams, specifically for tuning parameters in Adaptive Random Forest (ARF) and parameters for several drift detectors. We compare our approach against non-adaptive methods on both synthetic and real-world datasets. Empirical results show that our approach makes a compromises between predictive performance and resource consumption for ARF, and between true positives and false positives for drift detectors.