Research Articles (Computer Science)

Permanent URI for this collectionhttp://hdl.handle.net/2263/1695

Browse

Now showing 1 - 20 of 273

A multi-model ensemble for robust and consistent scoring of reflective essays
Baphumelele, Masikisiki; Marivate, Vukosi; Hlophe, Yvette Nkondo; Adam, Sumaiya (Institute of Electrical and Electronics Engineers, 2026-06-17)
Automated essay scoring has the potential to improve assessment efficiency, yet its adoption in medical education remains limited due to concerns regarding scoring reliability, robustness to linguistic variation, and the interpretability of automated feedback. This challenge is particularly evident in reflective writing assessment, where transparency and pedagogical relevance are essential. This study introduces LLAMPT (Linguistically-Aware Majority Prediction Transformer), an ensemble framework that combines multiple instruction-tuned large language models to generate assessment scores and explanatory rationales. The framework is trained and evaluated on a domain-specific dataset of 13,125 reflective essays written by medical students in South Africa, addressing an underrepresented context in AES research. Experimental results demonstrate strong agreement with expert human markers (Quadratic Weighted Kappa = 0.963). Robustness analysis using progressively paraphrased essays showed sustained scoring consistency across all transformation levels (QWK > 0.94), with mean score deviations below 0.125 rubric points and practical equivalence confirmed through TOST analysis. Interpretability was examined through attribution-based analyses and a mixed-methods evaluation involving clinical educators. While educators generally viewed generated justifications as coherent and contextually relevant, qualitative feedback highlighted limitations related to explanation faithfulness, stylistic nuance, and the continued need for human oversight. These findings demonstrate the potential of large language model ensembles for reflective writing assessment while highlighting important considerations for their responsible integration into medical education workflows.
Example-centric neuron labeling with weight-centric finalization : a comparative study
Van Heerden, Willem S. (Springer, 2026-02-12)
Self-organizing maps (SOMs) are unsupervised neural networks. SOMs use a map structure of neurons to model training data. Data science applications usually require neurons to be labeled. Three supervised neuron labeling techniques are commonly used, namely example-centric neuron labeling (ECNL), example-centric cluster labeling (ECCL), and weight-centric neuron labeling (WCNL). The ECNL algorithm produces high-quality labels but leaves some neurons unlabeled. Label finalization techniques complete the labeling of ECNL, ensuring high-quality labels that characterize the entire map. Recent work proposes a label finalization algorithm called example-centric neuron labeling with weight-centric finalization (ECNL-WCF) and demonstrates its feasibility. This article extends this previous work by providing a comprehensive analysis of ECNL-WCF in relation to existing supervised neuron labeling methods. An empirical investigation compared the performance of ECNL-WCF to ECNL, ECCL, and WCNL on data classification tasks. For each data set a 30-fold cross-validation was performed. Training set and test set classification errors and the percentages of unlabeled neurons were measured, and performance differences were confirmed using statistical hypothesis testing. Analysis showed that ECNL-WCF classified data as accurately as ECNL, and more accurately than both ECCL and WCNL. Additionally, ECNL-WCF fully labeled maps, which both ECNL and ECCL did not achieve. This study confirms that ECNL-WCF produces accurate labels, while fully labeling SOMs. These characteristics are attractive for exploratory data analysis, where a complete labeling makes maps more interpretable, but label accuracy is also important.
A comprehensive review of machine learning applications in cybersecurity : identifying gaps and advocating for cybersecurity auditing
Rananga, Ndaedzo; Venter, H.S. (Hein) (Springer, 2026-05-25)
The rapid growth of increasingly sophisticated and complex cyber threats has intensified interest in, and reliance on, artificial intelligence (AI), particularly machine learning (ML), within the cybersecurity landscape. ML has demonstrated robust potential to enhance cybersecurity capabilities, including threat detection, anomaly identification, predictive analytics, and automated response; however, practical ML implementation in cybersecurity remains in an early stage, often inconsistent, fragmented, and insufficiently developed. Consequently, ongoing research is essential to identify emerging developments, expand areas of inquiry, and propose improvements to existing approaches. This study provides a comprehensive review of recent ML applications in cybersecurity. Using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology, the study systematically evaluates the feasibility, effectiveness, and limitations of current approaches. The review reveals several critical gaps, including narrow application scopes, suboptimal algorithm performance in real-world environments, insufficient and imbalanced datasets, and inadequate integration with Security Information and Event Management (SIEM) and Intrusion Prevention Systems (IPS). Additional concerns include ethical dilemmas and governance challenges, all of which limit the operational reliability of ML-driven cybersecurity solutions. The findings emphasize the need to refine ML models, improve interoperability with existing security infrastructures, and strengthen evaluation frameworks. Importantly, the study advocates aligning modern ML-driven innovations with cybersecurity auditing, emphasizing cybersecurity audits' role in assessing ML readiness, validating control effectiveness, and promoting responsible, transparent, and risk-based adoption of ML technologies in cybersecurity environments.
A structure-based genetic programming generation constructive hyper-heuristic with transfer learning for combinatorial optimisation
Scheepers, Darius;; Pillay, Nelishia (Springer, 2025-12)
Generation constructive hyper-heuristics have proven to be very effective at creating construction heuristics for combinatorial optimization problems, with the heuristics derived by these hyper-heuristics often outperforming human-derived heuristics. Genetic programming has been predominantly used by generation constructive hyper-heuristics. This study investigates the effectiveness of two emerging technologies in genetic programming, namely, transfer learning and structure-based genetic programming, in genetic programming constructive generation hyper-heuristics. The study investigates whether structure-based genetic programming in generation constructive hyper-heuristics with (SBGP-HH-TL) and without transfer learning (SBGP-HH). The hyper-heuristics were evaluated on three problem domains, namely, the examination timetabling problem, the one-dimensional bin packing problem and the capacitated vehicle routing problem. Both SBGP-HH-TL and SBGP-HH outperformed the generational hyper-heuristic employing canonical genetic programming (CGP-HH) on a majority of the problem instances for the three problem domains, with SBGP-HH-TL outperforming SBGP-HH. Hence, the study has revealed that both transfer learning and structure-based genetic programming have resulted in performance improvements in genetic programming generation constructive hyper-heuristics for combinatorial optimization.
Cross-lingual embedding methods and applications : a systematic review for low-resourced scenarios
Sindane, Thapelo Andrew; Marivate, Vukosi; Modupe, Abiodun (Elsevier, 2025-09)
The field of Natural Language Processing (NLP) has achieved significant success in various areas, such as developing large-scale datasets, algorithmic complexity, optimized computing capabilities, refined individual and community expertise, and more, particularly in languages such as English, French, and Spanish. However, such global north unilateral strides have inadvertently created a substantial representation bias towards many languages categorized as low-resourced languages, with the majority being African languages. As a result, rudimentary resources such as stopwords, lemmatizers, stemmers, and word embeddings, as well as advanced multilingual transformer-based models remain under-developed for these languages. Compounding these circumstances is the lack of insights surrounding the development of these resources in the low-resourced context (e.g., how to develop embeddings for morphologically rich languages). Looking back, research priorities aiming to create these resources, largely motivated by the high cost attached to remedying these issues shifted, leading to the rise of alternative methods such as cross-lingual transfer learning (CLTL). CLTL involves transferring domain knowledge gained from supervised training to a domain with limited supervision signals. This study conducts a systematic literature review of CLTL techniques, in the context of cross-lingual models and embeddings, looking at their mathematical foundations, application domains, evaluation metrics, languages covered, and the latest developments. The findings of this study offer valuable insights into the present scenario of CLTL techniques, identifying areas for future research and development to advance cross-lingual natural language processing applications specifically in low-resourced settings.
Digital forensic readiness to mitigate insider threats in the SaaS cloud environment
Shoderu, Gabriel Olanrewaju Olumayowa; Baror, Stacey Omeleze; Modupe, Abiodun; Makura, Sheunesu M. (AI Society & STMIK Indonesia, 2025)
Insider threats continue to pose significant risks in Software-as-a-Service (SaaS) environments, where legitimate users hold varying levels of access and control. Existing mitigation measures remain largely reactive, focusing on postincident investigation and evidence recovery, which often result in delayed detection and incomplete forensics. A proactive and forensically sound approach is therefore required to identify and contain insider activity before major compromise occurs. This paper presents the Digital Forensic Readiness to Bust Insider Threats (DFR-BUST) model, a framework that embeds forensic readiness principles within SaaS environments to enable early detection, secure evidence capture, and legally defensible investigations. The model is aligned with the ISO/IEC 27043 digital investigation process, operationalising its readiness, acquisitive, and concurrent process classes. The model was evaluated using an experimental setup based on publicly available insider-threat datasets to demonstrate its readiness and detection capability. The evaluation confirmed that the proposed architecture supports proactive evidence generation, integrity verification, and traceable anomaly detection within a controlled environment. Unlike conventional reactive approaches, DFR-BUST provides a proactive, evidence-centric mechanism that enhances both detection accuracy and forensic admissibility. Its modular design ensures adaptability across cloud platforms while maintaining compliance with international forensic investigation standards. Overall, this work bridges the gap between intelligent analytics and digital forensic readiness. By ensuring that insider detection outputs are accompanied by verified, admissible evidence, the framework contributes a practical foundation for developing forensic-aware, cloud-based security systems.
Sentiment analysis using unsupervised learning for local government elections in South Africa
Matloga, Mokgadi Penelope; Marivate, Vukosi; Olaleye, Kayode (Department for E-Governance and Administration, 2025)
This study examines public sentiment during the 2021 South African local government election campaign by analysing Twitter posts. The research uses advanced techniques such as fine-tuned RoBERTa model, VADER, and TextBlob to assess the sentiments of tweets about four political parties, addressing the difficulties of understanding political sentiment on social media. The research also distinguishes tweets from real human users and those from chatbots, employing the K-Means method to detect suspicious activity. To gain deeper insights into the analysis, OpenAI GPT is employed for dataset labelling and managing class imbalance. The results show that sentiment varied significantly over time, with the fine-tuned RoBERTa model providing the most accurate analysis. The results further indicated that most tweets came from real human users, with a small number from bots, which tended to be negative. The findings offer useful insights for shaping political campaigns based on public sentiment trends.
A few-shot learning approach for a multilingual agro-information question answering system
Banda, Fiskani Ella; Marivate, Vukosi; Nakatumba-Nabende, Joyce (Wiley, 2025-04)
Across numerous households in Sub-Saharan Africa, agriculture plays a crucial role. One solution that can effectively bridge the support gap for farmers in the local community is a question–answer system based on agricultural expertise and agro-information. The more recent advancements in question answering research involve the use of large language models that are trained on an extensive amount of data. Due to this, conventional fine-tuning approaches have demonstrated a significant decline in performance when using a significantly smaller amount of data. One proposed alternative to address this decline is to use prompt-based fine-tuning, which allows the model to be fine-tuned with only a few examples, thus addressing the disparities between the objectives of pretraining and fine-tuning. Extensive research has been done on these methods, specifically on text classification and not question answering. In this research, our objective was to study the feasibility of recent few-shot learning approaches such as FewshotQA and Null-prompting for domain-specific agricultural data in four South African languages. We first explored creating a cross-lingual domain-specific extractive question answering dataset through an automated approach using the GPT model. Through exploratory data analysis, the GPT model was able to create a dataset, which requires minor improvements. We then evaluated the overall performance of the different approaches and investigated the effects of adapting these approaches to suit the new dataset. Results show these methods effectively capture semantic relationships and domain-specific terminology but exhibit limitations, including potential biases in automated annotation and plateauing F1 scores. This highlights the need for hybrid approaches that combine artificial intelligence and human supervision. Beyond academic insights, this study has practical significance for industry, demonstrating how prompt-based methods can help tailor AI models to specific use cases in low-resource settings.
On the classification of questioned PDF documents — attributing PDF documents to the tools that created them
Olivier, Martin S. (Elsevier, 2026-06)
It is often useful in forensic science to attribute an object to the tool that created it. In traditional questioned documents, it may be useful to determine which typewriter was used to produce them. In ballistics, it may be useful to determine which firearm fired a round found at a crime scene. We suggest that it is similarly useful to determine which tool was used to create a digital document. The current paper addresses the challenge of determining which tool created a PDF document. The first challenge is to show that a given tool consistently leaves marks in a PDF file that may be different from marks left by other tools. This is the first problem that the current paper addresses. It identifies potential toolmarks and then verifies that such marks occur consistently in a large dataset. The dataset used is the 1000 .gov PDF dataset provided by the US Library of Congress. Attribution of a creation tool aligns with the definition of classification provided by Inman and Rudin; they define a class as a set of objects that share a common origin. The paper extends the verification of consistency, by using the identified toolmarks to divide the dataset into classes with remarkable success. HIGHLIGHTS • A large number of immaterial aspects of PDF document representation are not (fully) standardised. • The manner in which a tool generates an aspect of PDF representation may be considered a toolmark. • Tools that create PDF documents leave consistent toolmarks in those documents. • Toolmarks may depend on tool configuration. • Toolmarks have the potential to identify the tool used to create the PDF file.
Exploring alternative teaching and learning platforms in post-COVID-19
Umera, Ozioma R.; Baror, Stacey Omeleze (Foundation of Advanced Science Education, 2025-06-10)
Learning for undergraduate students at the National Open University of Nigeria (NOUN), Gwagwalada study center, has significantly shifted toward e-learning since the COVID-19 pandemic. This study investigates how NOUN students use e-learning platforms and the challenges they face. Using a survey design with qualitative methods, data from 364 participants were collected through questionnaires and interviews. Guided by the Technology Acceptance Theory, the study finds that students predominantly use emails, LMS, video conferencing, and social media for academic activities. Key benefits include flexibility, convenience, equitable access, and enhanced critical thinking. However, infrastructure and policy gaps persist. The study highlights the importance of ongoing improvements in e-learning systems and recommends that NOUN regularly conduct faculty training and digital literacy workshops to enhance participation and reduce barriers.
Scoping review of artificial intelligence via mobile technology and social media for health in Africa
Baichoo , Shakuntala; Oladeji, Olubusola; Villareal, Leanne; Diakabana, Huguette; Okekunle, Akinkunmi Paul; Marivate, Vukosi; Kaggwa, Fred; Nsoesie, Elaine O. (Nature Research, 2025-12-20)
The combination of mobile technologies and social media with Artificial Intelligence (AI) opens new opportunities for multi-modal data generation, analysis, and inference for various health applications. To investigate how these tools are being used for health applications in Africa, we conduct a scoping review following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) approach. We screen 469 articles and synthesize 116. We include 29 studies documenting the use of a broad range of advanced and straightforward machine-learning techniques to study infectious and chronic diseases such as COVID-19 (4 studies, 13.8%), malaria (5, 17.2%), and cervical cancer (2, 6.9%). Countries with high internet and mobile phone penetration have higher representation. Based on identified gaps, we make research and policy recommendations to enhance the contribution of these tools in advancing health in Africa. These include investing in studies on chronic diseases and implementing frameworks to address geographic inequity.
Privacy vs. utility in federated learning : an experimental analysis of noise injection techniques
Leope, Neo R.; Eloff, Jan H.P.; Dlamini, Moses Thandokuhle (Institute of Electrical and Electronics Engineers, 2025-11-20)
Federated Learning (FL) enables decentralized model training, while maintaining the privacy of the underlying individual datasets. Therefore, FL can resolve some intrinsically privacy-sensitive challenges in domains, such as healthcare and finance. However, privacy preservation usually comes with a trade-off on the usefulness (i.e., utility) of the information. The research problem is how to optimize this inversely proportional trade-off balance between privacy and utility. This study uses an experimental comparative analysis, in a synthetic healthcare setting, of different noise types (i.e., Gaussian, Laplacian, Poisson, Uniform, and Exponential) injected on the client side at the input-feature level prior to local training to enhance privacy in FL. We explore the impact of these noise types on the privacy–utility trade-off in FL data. The findings indicate that although Laplacian, Poisson, and Exponential types of noise provides stronger obfuscation which often comes at the cost of utility. This confirms and amplifies the trade-off in maintaining the usefulness of the data against its privacy. More importantly, the findings also show that Gaussian noise generally offers the best trade-off between privacy and utility on this task, suggesting a practical default for privacy-aware FL in healthcare-like environments.
A swarm intelligence-based hybrid metaheuristic with tabu search for the quadratic assignment problem
Panwar, Karuna; Rajwar, Kanchan; Deep, Kusum; Cho, Sung-Bae (Springer, 2026-02-04)
The Grey Wolf Optimizer (GWO), inspired by the hunting behavior of grey wolves, is an effective swarm intelligence-based algorithm increasingly recognized for solving NP-hard problems. The Quadratic Assignment Problem (QAP), known for its complexity and widespread industrial applications, presents a significant challenge in combinatorial optimization. This paper introduces a novel discrete variant of GWO for QAP, the Hybrid Grey Wolf Optimizer (HGWO), which integrates an enhanced Tabu Search (TS) to improve GWO’s effectiveness in solving the QAP. This enhanced TS is employed to refine the exploitation phase by focusing on promising areas identified by GWO. Due to the combinatorial nature of QAP, the outcomes of classical GWO are transformed into discrete values using the largest real value mapping technique. In our computational experiments across all 134 QAPLIB benchmark instances, HGWO achieved the best-known solutions for 110 instances. It maintains an impressively low average deviation of 0.20%, demonstrating high accuracy and robustness. Comparative analysis with established algorithms like Genetic Algorithm, Bat Algorithm, and Whale Optimization Algorithm demonstrates that HGWO surpasses most competing methods. Rigorous statistical tests, including the Friedman nonparametric test and the Wilcoxon signed-rank test, validate these results, underscoring HGWO’s potential as a powerful tool for QAP and indicating fruitful directions for future research in combinatorial optimization strategies.
Blockchain forensics and regulatory technology for crypto tax compliance : a state-of-the-art review and emerging directions in the South African context
Ramazhamba, Pardon Takalani; Venter, H.S. (Hein) (MDPI, 2026-01-13)
The rise in Blockchain-based digital assets has transformed the financial ecosystems, which has also created complex governance and taxation challenges. The pseudonymous and cross-border nature of crypto transactions undermines traditional tax enforcement, leaving regulators such as the South African Revenue Service (SARS) reliant on voluntary disclosures with limited verification mechanisms, while existing Blockchain forensic tools and regulatory technologies (RegTechs) have advanced in anti-money laundering and institutional compliance, their integration into issues related to taxpayer compliance and locally adapted solutions remains underdeveloped. Therefore, this study conducts a state-of-the-art review of Blockchain forensics, RegTech innovations, and crypto tax frameworks to identify gaps in the crypto tax compliance space. Then, this study builds on these insights and proposes a conceptual model that integrates digital forensics, cost basis automation aligned with SARS rules, wallet interaction mapping, and non-fungible tokens (NFTs) as verifiable audit anchors. The contributions of this study are threefold: theoretically, which reconceptualise the adoption of Blockchain forensics as a proactive compliance mechanism; practically, it conceptualises a locally adapted proof-of-concept for diverse transaction types, including DeFi and NFTs; and lastly, innovatively, which introduces NFTs to enhance auditability, trust, and transparency in digital tax compliance.
An anti-sheriff cybersecurity audit model : from compliance checklists to intelligence-supported cyber risk auditing
Rananga, Ndaedzo; Venter, H.S. (Hein) (MDPI, 2026-03)
The increasing adoption of data-driven techniques in cybersecurity has introduced new opportunities to enhance detection, response, and automation capabilities within the cybersecurity ecosystem; however, cybersecurity auditing remains constrained by traditional compliance-oriented approaches that rely profoundly on binary, checklist-based evaluations. Such approaches often reinforce a policing or “sheriff-style” perception of auditing, emphasizing enforcement rather than enablement, risk insight, and organizational improvement. Of primary concern is that the “sheriff-style” cybersecurity audit approach often fails to accurately portray the true state of an organization’s cybersecurity posture, often providing a misleading sense of assurance based solely on formal compliance and controls existence. This study proposes an Anti-Sheriff Cybersecurity Audit Model, that moves beyond cybersecurity control checklists, by integrating intelligence-informed risk assessments with structured human judgment to support a more robust, adaptive, and risk-oriented auditing process. Grounded in design science research (DSR), the proposed approach combines conventional binary compliance verification with intelligence-derived risk indicators and governance-based maturity assessments to evaluate cybersecurity controls across technical, operational, and organizational dimensions. The approach aligns with established standards and frameworks, including International Organization for Standardization and the International Electrotechnical Commission (ISO/IEC) 27001, the National Institute of Standards and Technology (NIST), and the Center for Internet Security (CIS) benchmarks, while extending their application beyond static compliance validation. A fictional case study is used to demonstrate the model’s applicability and to illustrate how hybrid scoring can reveal residual risk not captured by conventional cybersecurity audits. The findings indicate that combining intelligence-informed analytics with structured human judgment enhances audit depth, interpretability, and business relevance. The proposed approach, therefore, provides a foundation for evolving cybersecurity auditing from just periodic compliance assessments, toward a continuous, risk-informed, and governance-aligned assurance system.
Advancements in maize yield estimation : a comprehensive review of methods and models
Hove, Kudakwashe; Nyamugure, Philimon; Mdlongwa, Precious; Dube, Timothy; Nyathi, Thambo; Awala, Simon Kamwele (Springer, 2025-12-23)
Accurate and timely estimation of maize yield is crucial for ensuring food security, optimizing resource utilization, and informing agricultural policy. However, current yield estimation methods often encounter significant limitations, such as low spatial resolution, dependence on sparse ground-truth data, poor model generalizability across diverse agroecological zones, and challenges in integrating heterogeneous data sources. Although numerous techniques have been developed, ranging from traditional field-based measurements to advanced remote sensing and machine learning methods, a comprehensive synthesis that critically evaluates these approaches and explores their convergence is still lacking. This review addresses this gap by providing a systematic overview of recent advances in maize yield estimation, with a focus on remote sensing technologies, machine learning algorithms, and hybrid crop modeling frameworks. It examines the strengths and limitations of various methodologies, including UAV- and satellite-based imaging, hyperspectral and LiDAR sensing, regression and ensemble learning, and long-read sequencing. Additionally, the review explores the role of emerging technologies such as IoT, cloud computing, and blockchain in enhancing data collection, processing, and traceability. By identifying key challenges such as environmental variability, data scarcity, and model interpretability, and highlighting opportunities for methodological integration, this review offers a roadmap for future research and development. It argues that the convergence of digital agriculture tools and robust modeling strategies holds significant promise for improving maize yield estimation accuracy, scalability, and applicability. These advancements have far-reaching implications for sustainable agriculture, climate resilience, and global food security.
ZASCA-sum : a dataset of the South Africa supreme courts of appeal judgments and media summaries for legal documents summarization research
Adulmumin, Idris; Marivate, Vukosi (Elsevier, 2025-06)
This paper presents ZASCA-Sum, a novel dataset comprising judgments from the South Africa Supreme Court of Appeal and their manually curated media summaries. The dataset, collected from the court's official website, includes 4171 judgments, of which 2118 have summary pairs. The judgments and summaries have been extracted and prepared to support legal document summarization tasks across supervised, semi-supervised, and unsupervised settings. This paper provides a detailed description of the dataset, covering the data collection process, timeline, processing, and potential applications in the field. We provide the token-count distribution and analysis of the judgments and summaries that can be accommodated off-the-shelf by current summarization models with the largest input token size. The dataset, split into training, validation, and test sets, is made publicly available to encourage research in legal summarization. In addition to document summarization, researchers can use this data to localize English-centric models to support the South African dialect.
A selection perturbative hyper-heuristic for neural architecture search
De Clercq, Johannes; Pillay, Nelishia (Elsevier, 2026-03)
Neural architecture search explores the architecture space, referred to as the design spaces, to find an architecture that produces good results. Various approaches, such as genetic algorithms, are usually used to explore this space. This study investigates exploring an alternative space, namely, the heuristic space using a hyper-heuristic to indirectly explore the design space. The study introduces the concept of a NAS operator space (NOS). A single point selection perturbative hyper-heuristic (SPHH-NAS) explores a heuristic space that maps to the NOS which then maps to the design space. A choice function is used for heuristic selection and the Adaptive Improvement Limited Target Acceptance (AILTA) for move acceptance. It is anticipated that indirectly searching the design space will facilitate reaching areas of the search space that could not be reached by searching the space directly. SPHH-NAS was evaluated on three NAS benchmark sets, namely, NAS-101, NAS-201 and NAS-301. In addition to this the approach is evaluated on two real-world datasets. SPHH-NAS was found to outperform majority of the previous approaches used to solve these problems. In addition to this SPHH-NAS resulted in a reduction in computational cost. HIGHLIGHTS • This is the first study using a selection perturbative hyper-heuristics for neural architecture search. • The selection perturbative hyper-heuristic produces good results for NAS. • The selection perturbative reduces computational cost for NAS.
Fine-tuning a sentence transformer for DNA
Mokoatle, Mpho; Marivate, Vukosi; Mapiye, Darlington; Bornman, Maria S. (Riana); Hayes, Vanessa M. (BioMed Central, 2025-10)
BACKGROUND : Sentence-transformers is a library that provides easy methods for generating embeddings for sentences, paragraphs, and images. Sentiment analysis, retrieval, and clustering are among the applications made possible by the embedding of texts in a vector space where similar texts are located close to one another. This study fine-tunes a sentence transformer model designed for natural language on DNA text and subsequently evaluates it across eight benchmark tasks. The objective is to assess the efficacy of this transformer in comparison to domain-specific DNA transformers, like DNABERT and the Nucleotide transformer. RESULTS : The findings indicated that the refined proposed model generated DNA embeddings that exceeded DNABERT in multiple tasks. However, the proposed model was not superior to the nucleotide transformer in terms of raw classification accuracy. The nucleotide transformer excelled in most tasks; but, this superiority incurred significant computing expenses, rendering it impractical for resource-constrained environments such as low- and middle-income countries (LMICs). The nucleotide transformer also performed worse on retrieval tasks and embedding extraction time. Consequently, the proposed model presents a viable option that balances performance and accuracy.
Early detection of Phytophthora root rot in Eucalyptus using hyperspectral reflectance and machine learning
Esterhuizen, Hendrik J.; Slippers, Bernard; Bosman, Anna Sergeevna; Roux, Jolanda; Jones, Wayne; Bose, Tanay; Hammerbacher, Almuth (Elsevier, 2025-10)
The rising prevalence of Phytophthora diseases in forests highlights the need for rapid, non-invasive detection methods. Early-stage root infections are difficult to detect due to the absence of visible above-ground symptoms, while current diagnostics remain slow and invasive. This study investigated whether hyperspectral leaf reflectance could detect root rot caused by Phytophthora alticola in Eucalyptus benthamii. Nineteen commercially planted families were inoculated, and leaf spectra were collected using an ASD FieldSpec 4 sensor. A machine learning pipeline was developed to identify diagnostic spectral signals. Key wavelengths were identified using permutation importance, a genetic algorithm, and self-attention network (SAN) scores. Spectral signals linked to root rot revealed that infection was correlated with leaf pigment accumulation and moisture stress. Three algorithms, random forest (RF), support vector machine (SVM), and SAN, were trained on hyperspectral data to predict P. alticola infection. The SAN achieved 97 % accuracy on a reduced dataset, which included the diagnostic wavelengths from the feature selection step, surpassing the RF (96 %) and SVM (94 %) models. This study demonstrates hyperspectral sensing as an effective tool for detecting Phytophthora root rot using spectra from the foliage and highlights the application of advanced machine learning techniques for plant disease classification. HIGHLIGHTS • Hyperspectral sensing detects Phytophthora root rot before symptoms appear. • SAN model achieved 97 % accuracy using selected wavelengths from leaf spectra. • Key wavelengths correlated with pigment shifts and moisture stress in leaves. • Machine learning identified spectral markers for early disease detection. • Vegetation indices NDNI and MSI are strongly linked to infection status.

Browse

Recent Submissions