We predict weekly telephony contact for Centrelink customers using longitudinal administrative data aligned to a weekly grid, comparing statistical, machine learning, and deep learning models on 889,879 customers (150 features; 3.65M observations after class balancing; 5-fold CV). Feature-engineering regimes spanned from quick encodings to contact-history windows of 1, 4, 8, and 12 weeks; an 8-week history yielded the best trade-off, with diminishing returns beyond 8 weeks. Across 12 models, a Multilayer Perceptron attained the strongest performance on held-out data (AUC ≈ 0.84; accuracy ≈ 0.79). One-factor-at-a-time hyperparameter searches produced modest, non-monotonic gains, underscoring that additional tuning effort does not necessarily translate to uplift in model performance. To guide modelling investment and effort decisions, we introduced a cost-based budgeting equation associated with treatment applications of the predictive model. This framework approach encourages monetary costings to be considered alongside predictive performance for model development.
Software vulnerabilities are referred to as weaknesses in the source code. Hackers exploit these weaknesses to perform malicious actions, including accessing sensitive data and injecting a computer virus to hijack the computer system. Identifying these vulnerabilities is challenging. Even a perfectly functional program may have hidden vulnerable patterns. Contrastingly, locating these patterns manually by cybersecurity experts is onerous and time-consuming. Existing research uses deep learning algorithms to automate the process of finding vulnerable patterns. However, extracting relevant features from the source code is challenging, as source code can exceed hundreds of lines. Existing researchers either extract individual functions in a program or encapsulate the entire program, leading to under- or over-representation, respectively. A recent notion of using program slices has emerged, which only retains the most likely vulnerabilities causing statements. However, no significant research has been conducted to identify the most significant characteristics that are likely to cause vulnerabilities. Therefore, in this research, we investigate statistical heuristics to dynamically determine the most representative vulnerability characteristics and propose a deep neural network based on RoBERTa embeddings, which is fine-tuned on CodeBERT using multi-sample dropout to enhance generalizability. Our experimental results on real-world software source databases show up to 37
In deep learning-based image classification, it is necessary to achieve high generalization performance while suppressing overfitting in situations when training data are limited. Mixup, a data augmentation using linear interpolation in the input space, is effective in smoothing the decision boundary, but it can hinder the model’s ability to learn fine-grained, class-specific details in the later training stages. On the other hand, augmentation by First-Order Manifold Augmentation (FOMA) utilizes local geometry but is designed for regression, making its direct application to classification difficult. We propose a two-phase augmentation method that changes its strategy according to the phase of training. In Phase 1, standard Mixup is used for global interpolation to smooth the class boundary. In Phase 2, we apply Local-FOMA, which is designed for classification, to estimate principal component directions based on local neighborhoods in the feature space and perform data augmentation while preserving the manifold structure. The weights of the loss function are linearly transitioned during the learning process, enabling a smooth switch from Mixup to Local-FOMA. On CIFAR-100 and CIFAR-10, our method outperformed strong Mixup variants. Under distribution shifts (CIFAR-C), it also improves robustness and was particularly effective on CIFAR-100-C, suggesting that the method is particularly advantageous in complex scenarios involving datasets with many classes. These results demonstrate that Mixup-FOMA provides an effective strategy to enhance both classification performance and robustness by gradually combining global generalization with local refinement.
Forecasting stock market indices is challenging due to the dynamic and non-stationary nature of financial data. Traditional machine learning methods often struggle to capture these complexities, limiting their accuracy. In this paper, we propose two novel dynamic meta-learning ensembles—DME-PC and DME-HC, to address the critical challenge of ensemble member selection and adapting the prediction to each predicted data point, to enhance accuracy. To select ensemble members, DME-PC employs pairwise comparisons with statistical testing while DME-HC uses hierarchical clustering. To adapt the ensemble to the characteristics of the predicted data points, DME-PC and DME-HC combine dynamically the ensemble member predictions based on their performance on recent data and predicted future performance. Experiments on four financial datasets for 10 years and comparison with statistical, machine learning and deep learning methods showed the superior performance of the proposed dynamic ensembles. The best results were obtained with DME-HC (EWA), which utilizes hierarchical clustering for ensemble member selection and exponentially weighted average on most recent errors. Our results highlight the potential of the proposed dynamic ensembles to effectively capture temporal patterns in financial data and support better investment decisions.
Multi-omics data, encompassing heterogeneous sources such as genomics, transcriptomics, and proteomics, has become a powerful resource for disease classification in bioinformatics. Fusion mechanisms combine features from multiple modalities into a single joint representation for predictive models to improve classification performance. However, while most studies focus on accuracy, model calibration - how well predicted probabilities match actual outcomes - remains underexplored in the multi-omics setting. Poor calibration can make even accurate models overconfident and unreliable. This study investigates the impact of three fusion strategies, early, intermediate and late, on three state-of-the-art post-hoc calibration techniques. We further introduce dynamic scaling, a regression-based calibration method that estimates instance-specific temperature parameters for more adaptive calibration than dataset-level techniques. Experiments on four benchmark omics datasets evaluate both classification and calibration metrics. Results show that data quality and class imbalance strongly influence performance, and while dynamic scaling often achieves the best calibration, challenges persist for small and imbalanced datasets.
Urban traffic prediction is hindered by heterogeneous sensor configurations and complex wave-like traffic flow dynamics (e.g., congestion waves, stop-and-go oscillations) that conventional neural networks struggle to model effectively. We propose WaveFSL, a novel wave physics-inspired few-shot learning framework that unifies wave interference theory with adaptive neural modules to address these challenges. WaveFSL integrates five key components: (1) a Dynamic Input Projection (DIP) for handling variable-dimensional input through learnable dimension-aware projection; (2) a Traffic Wave Generator (TWG) synthesising parameterised wave components (with amplitude, effective frequency, and phase); (3) a Wave-Constrained Interference (WCI) explicitly modelling congestion propagation via coupled wave superposition; (4) a Wave-Aware Spectral Attention (WASA) for multiscale spectral analysis through resonance scoring and frequency-band decomposition; and (5) a Few-Shot Adaptation with optimal kernels (FSAK) enabling rapid domain adaption and transfer via prototype-based conditioning. By unifying wave interference theory with adaptive neural learning, WaveFSL achieves state-of-the-art performance across four real-world datasets, outperforms baselines (3.2–8.5 https://github.com/afofanah/WaveFSL .
This study investigates how document structure and preprocessing strategies influence the performance of Retrieval-Augmented Generation (RAG) pipelines in domain-specific question answering, using university curriculum handbooks as a case study. We evaluate multiple data representations, varying in chunking method, punctuation level, and structural format (HTML, plain text, JSON), within a privacy-preserving pipeline that combines semantic vector indexing with a lightweight, locally deployed LLM. A benchmark of curriculum-related queries, covering both simple lookups and multi-hop reasoning, was used to assess performance under controlled retrieval conditions. Results show that semantically coherent chunking, manual or LangChain-based, substantially improves accuracy, with gains exceeding 25
Knowledge Graphs (KG) form the backbone of many knowledge dependent applications, such as search engines and digital personal assistants. When constructing a KG, data can be manually curated by experts, contributed by volunteers, automatically extracted using hand-crafted or learned rules, or generated from unstructured text via machine learning techniques. Regardless of the approach, anomalies are inevitable, as no data source is perfect. To address this, we propose ETCOD, an embedding-based anomaly detection approach for KG validation and quality enhancement, combined with Large Language Models (LLM) for explanation and verification. First, we generate semantic embeddings of triples in order to capture entity and relation similarities. Next, we perform pattern mining to identify anomalous triples. Finally, detected anomalies are forwarded to an LLM to provide human-understandable explanations and reasoning. We conducted experiments on real-world KGs, including YAGO-1 and YAGO-4.5, using ChatGPT-4o and Gemini for explanation. Our analysis highlights which types of anomalies are most effectively explained by each model and where they tend to fall short. The results demonstrate that embedding-based detection is effective in identifying anomalies, while LLMs enhance interpretability by providing context-aware explanations.
Missing data is a problem commonly seen in most if not all real-world applications. Particularly for water quality monitoring systems, which are commonly plagued by sensor faults or network errors, missing or erroneous data pose a significant challenge in extracting accurate and meaningful insights. In this work, we investigate the problem of missing information in time-series data and propose a new method DISC - Data Imputation with Seasonality and Causality - which uses the concepts of seasonal decomposition and causal discovery to improve contextual accuracy of the imputations for time-series. DISC operates in two stages. First, it builds a causal relational graph representing inter-feature dependencies and uses this graph to impute missing values by adjusting estimates to the nearest-neighbour hourly data points. Second, it learns yearly, monthly, and daily seasonal patterns at an hourly resolution and imputes the remaining gaps. In scenarios where seasonal decomposition fails to fully resolve gaps, causal discovery exploits dependencies among time-series features to generate reference points that enhance the completeness of the seasonal pattern. The proposed method has been evaluated using real-world data from the Murray-Darling Basin and compared with multiple existing machine learning methods. The results validate the effectiveness of DISC, enabling accurate imputation of 14 consecutive days of missing hourly data with an R-squared of 80
Ranking individuals by relevance is central to many decision-making systems, such as hiring. A common approach is top-k ranking, where only the best k individuals are selected. However, social biases in data can make automated ranking systems discriminatory and affect individuals’ lives. In line with the AI Act, fair top-k ranking has recently gained attention. In this work, we focus on positional fairness in top-k rankings, which ensures equal exposure of groups across ranking positions. We formulate this as a constrained optimization problem that balances positional fairness with ranking quality. We propose two novel exact algorithms with optimal guarantees, along with an efficient greedy approximation. All methods support non-binary and multiple protected attributes as often present in real-world. Experiments on synthetic and real-world datasets show that our methods outperform recent approaches in both positional fairness and ranking quality.
Precise categorization of sheep parasite eggs is essential for improved veterinary diagnostics, automated monitoring, early disease detection, and efficient farm management in modern agriculture. However, most existing deep learning models struggle to achieve both rich feature extraction and fast inference, often excelling at one while sacrificing the other. To address this limitation, we propose a DenseNet–YOLOv8 hybrid model that combines the dense connectivity and strong feature propagation of DenseNet with the anchor-free detection and high-speed inference of YOLOv8. The framework was trained and evaluated on a sheep parasite egg dataset under challenging conditions such as occlusion, illumination variation, and background clutter. Experimental results show that the hybrid model outperforms the individual DenseNet and YOLOv8 baselines across all tested metrics, achieving 96.8
Efficient retrieval of wildfire-related imagery from large-scale remote sensing and aerial datasets is essential for early detection, monitoring, and disaster response. However, the limited availability of annotated data severely restricts the effectiveness of conventional supervised approaches. To address this challenge, we propose a parameter-free self-distillation framework that leverages language models and unlabeled imagery to enhance wildfire image retrieval. Our method first employs large language models (LLMs) to generate domain-specific textual descriptions of wildfire phenomena, covering diverse visual cues such as fire spread, smoke plumes, vegetation conditions, and multi-perspective observations. These descriptions are used to construct a text-based classifier that transfers semantic knowledge into the visual domain. Through a self-distillation process, the classifier produces pseudo-labels for unlabeled wildfire imagery, which are then used to iteratively refine the vision encoder in a parameter-efficient manner. Experiments on both satellite remote sensing and aerial datasets demonstrate substantial improvements over zero-shot baselines, yielding more accurate and robust retrieval performance under class imbalance, heterogeneous imaging conditions, and cross-dataset generalization. This framework highlights the potential of combining LLM-driven semantic enrichment with self-distillation for scalable, annotation-free wildfire monitoring and supports broader applications in label-scarce environmental image analysis.
Federated Learning (FL) has emerged as a promising paradigm for collaborative model training without centralizing client data. However, most existing methods rely on single-objective optimization and heuristic aggregation strategies that neglect client-specific characteristics, resulting in performance degradation, unfair model behavior, and inefficient convergence under heterogeneous client settings. In this work, we propose FedMOAR, a Multi-Objective Adaptive Regularization strategy that jointly optimizes global model accuracy, client-level fairness, and communication efficiency. Unlike conventional FL approaches that apply uniform regularization or focus solely on minimizing global loss, FedMOAR dynamically adjusts its regularization coefficients based on model divergence, fairness penalties, and accuracy compensation. We evaluate FedMOAR on MNIST and NSL-KDD datasets using Dirichlet-based heterogeneous (non-IID) data partitions ( α = 0.1, 0.5, 1.0) that induce variability in data volume and class distributions under both partial and full client participation. Experimental results show that FedMOAR consistently outperforms baselines such as FedAvg, FedProx, FairFed, and FedVal. Specifically, it achieves higher F1-scores, lower min-max accuracy gap (MMAG), and improved Jain’s Fairness Index (JFI), while demonstrating up to 2.4 × speedup. Even in scenarios with comparable test accuracy, FedMOAR yields significantly better F1 and JFI values, and lower MMAG, confirming its effectiveness as a fair and efficient FL solution. These results highlight FedMOAR’s practical value in real-world deployments characterized by heterogeneous data and client diversity.
Recent advances in deep learning, particularly transformer-based models, have shown remarkable success in capturing both spatial and temporal dependencies in video and image data. Interestingly, financial time series data share analogous structures: asset returns at different time points form sequences, and the covariance matrix captures inter-asset relationships, similar to spatial correlations in images. Despite this analogy, existing financial forecasting models rarely leverage these spatio-temporal patterns to their full potential. This research addresses this gap by leveraging the Time-Space Transformer (TimeSformer) model, originally developed for video understanding, to analyze and predict financial time series data. Specifically, we transform the return features of portfolio stocks into image representations with a temporal structure. We experiment with four different TimeSformer architectures, employing both parametric and non-parametric loss functions to predict the subsequent image, which is then back-transformed into the corresponding returns and covariance matrices. Experimental results demonstrate that this approach greatly improves prediction accuracy, achieving lower RMSE for returns and reduced log-euclidean distance for covariance matrices compared to traditional methods.
Accurate and interpretable evaluation of vision language models (VLMs) is crucial for applications where accuracy and transparency are essential, such as automotive damage assessment, where structured outputs support reliable decision making. This paper presents CarDamageEval, a dual-layer evaluation framework designed to measure both the structural accuracy and semantic quality of VLM outputs. The framework enforces a predefined structured output format comprising explicit tuples of damage type, vehicle body part and severity level, enabling rigorous quantitative assessment through pair-matching metrics such as precision, recall and F1 score. Complementing this, semantic quality is assessed using the Holistic Description Score (HDS), which captures correctness, completeness, coherence and relevance. These two perspectives are unified through the hybrid CarDD Score, providing a balanced metric that rewards factual accuracy and descriptive clarity. To support structured evaluation, we compiled a dataset integrating public vehicle images, each annotated with bounding boxes and linked damage type, body part and severity labels. This baseline comparison further demonstrates the framework’s ability to distinguish model performance across configurations, highlighting the effectiveness of fine-tuning in generating accurate and structured damage descriptions. While developed for automotive assessment, the principles underlying CarDamageEval also apply to other structured vision language tasks, such as lesion detection in medical imaging or defect localisation in industrial inspection, making it a versatile and reusable evaluation standard.
Chemotherapy, including neoadjuvant chemotherapy administered prior to surgery, is a fundamental treatment strategy for breast cancer, yet patient responses are highly variable. Current chemotherapy planning primarily depends on a limited set of clinical and pathological factors, which often neglects the complex molecular heterogeneity inherent in individual tumours. Existing predictive models, while valuable, generally fail to address the essential clinical question: Which patients truly benefit from chemotherapy, and how should drug combinations be optimally tailored to individual patients? In response to this critical gap, we propose CTR (Causality-based Therapy Recommendation), a novel causal recommendation framework designed to personalise chemotherapy decisions in breast cancer treatment. CTR integrates causal inference methods, specifically causal trees, to estimate heterogeneous treatment effects and provide recommendations regarding whether chemotherapy would benefit individual patients, particularly in terms of achieving pathological complete response (pCR) and improved survival outcomes. Moreover, CTR extends its application by optimising combinations of chemotherapy agents, such as taxanes, anthracyclines, and anti-HER2 therapies, to enhance treatment efficacy. Evaluation of CTR on the DUKE and TransNEO datasets demonstrates significant improvements over existing state-of-the-art methods in terms of improved survival and recovery rates. CTR thus represents a promising step towards personalised precision oncology by enabling clinicians to make informed, data-driven chemotherapy decisions.
Spatial crowdsourcing (SC) platforms rely on predictive models to allocate tasks to mobile workers. However, real-world dynamics such as traffic, weather, and user behaviour cause frequent data and concept drift, which degrade model performance. Traditional drift detection methods are noise-sensitive, spatially agnostic, and computationally expensive. We present DriftSense, a hybrid incremental learning approach for real-time drift detection and adaptation in SC systems. DriftSense introduces three innovations: (i) spatially localised entropy-based drift detection, (ii) model-aware ADWIN (MA-ADWIN) that incorporates internal signals from Adaptive Hoeffding Trees (AHTs), and (iii) a false-signal filtering mechanism for robust adaptation. Experiments on real-world NYC Taxi and Yelp datasets, with injected abrupt, gradual, and mixed drifts, show that DriftSense achieves up to 25
Accurate forecasting of aboveground biomass (AGB) is essential for sustainable pasture management. Existing approaches, such as traditional machine learning (ML) and biophysical models, have limitations: ML models fail to effectively capture temporal dependencies and underlying biophysical processes, whereas biophysical models are difficult to scale due to their requirement for extensive calibration. Although hybrid frameworks that integrate biophysical model outputs into ML models show potential, they still cannot effectively exploit temporal information. To address these issues, we propose a novel framework that combines a multi-view Gate-Controlled LSTM (GC-LSTM) with the biophysical model Modvege for short-term AGB forecasting. Our GC-LSTM employs dual branches to model both climate data and Modvege-derived biophysical variables, adaptively balancing their contributions through a gating mechanism. We evaluate GC-LSTM against nine baselines, including Modvege and eight ML and deep learning models that incorporate biophysical variables using three independent yearly datasets (2021, 2022, and 2023) of 56 Australian paddocks. Results demonstrate that GC-LSTM consistently achieves higher accuracy, highlighting its robustness and practical utility for AGB prediction.
Accurate automated tree crown delineation (TCD) requires highly precise boundary segmentation, yet reduced-resolution architectures face limitations from decoder outputs at lower spatial resolutions. We propose BARE (Boundary-Aware with Resolution Enhancement), a simple architecture-preserving training strategy combining external full-resolution loss supervision with class weighting. BARE upsamples decoder output solely during training, maintaining inference efficiency while improving boundary precision. Through comprehensive evaluation on SegFormer, PSPNet, and SETR using the OAM-TCD dataset, we demonstrate that external full-resolution supervision universally benefits all tested architectures, achieving significant boundary quality improvements. We introduce B-IoU (Boundary-Intersection over Union) to TCD research, enabling rigorous boundary quality assessment. Our systematic evaluation reveals architecture-dependent optimization characteristics, providing actionable guidelines for practitioners seeking to enhance boundary precision in reduced-resolution segmentation architectures via training-only modifications. Code: https://github.com/attavit14203638/bare .
Accurate multiclass classification of dermoscopic images is critical for early detection of malignant skin lesions, yet existing high-performing deep learning approaches often require computationally heavy backbones, limiting deployment on mobile or point-of-care devices. This study proposes LightSkinNet, a lightweight convolutional neural network (CNN) integrating Dilated Convolutional Enhancement (DCE) blocks, the Convolutional Block Attention Module (CBAM), and Efficient Channel Attention (ECA) to capture multi-scale features and refine salient lesion patterns with minimal computational cost. A hybrid class-balancing pipeline combining mild undersampling and oversampling was applied to the training set, followed by stratified splitting to preserve natural validation/test distributions, and targeted on-the-fly augmentation to enhance generalization. LightSkinNet was benchmarked on the HAM10000 dataset using five-fold stratified cross-validation against baseline models. The proposed model achieved a Top-1 accuracy of 90.27%± 0.48 , Top-3 accuracy of 97.86%± 0.22 , macro F1-score of 89.12%± 0.53 , and Cohen’s Kappa of 0.872± 0.006 , outperforming all baselines with statistically significant improvements ( p<0.05 ) while maintaining a low parameter count. Gradient-Weighted Class Activation Mapping (Grad-CAM) showed that LightSkinNet focuses on key dermoscopic features, enhancing transparency in AI-assisted diagnosis. These results demonstrate that LightSkinNet delivers an effective accuracy–efficiency trade-off suitable for real-time clinical decision support on resource-constrained platforms.
