With more and more institutions adopting online delivery of education, the accompanying need for assessing students online is increasing. However, the tendency of some candidates to engage in fraudulent acts to gain undue advantage over other candidates undermine the integrity of the online testing process. It is a major challenge faced by test administering institutions. The most widespread fraudulent act is impersonation. Existing studies primarily focus on the detection of this particular fraudulent act. However, there are other fraudulent acts as well, resorted to by cheating candidates. Methods of detecting these other acts have not been sufficiently studied. The objective of the study presented in this paper was to address this gap through the extension of a non-obtrusive authentication system intended for the detection of impersonations during online tests, developed and tested by the authors, to detect fraudulent acts other than impersonation. The system used keystroke dynamics of the candidates for authentication. This study was particularly aimed at examining the capability of the already tested imposter detection system to detect other fraudulent acts as well. Experiments were conducted using 37 volunteers roleplaying as online test candidates amidst simulated scenarios of fraudulent test taking other than impersonation and their keystroke data were acquired. These data were used to play back the test taking offline, while exercising continuous authentication using the imposter detection system. Quantitative analysis of the resulting data indicates that the imposter detection system can detect the fraudulent acts of look up and collaboration with accuracies of 51.7% and 59.9% respectively, which are dependable but marginally lower than the accuracy of imposter detection. This implies that the system is inherently capable of detecting fraudulent acts other than impersonation, although with lesser accuracy. Plausible reasons for the lower accuracy in detecting other fraudulent acts and ways of improving it are discussed.

Hate speech is a major issue accompanying the exponential increase of social networking sites. Twitter is a microblogging based platform that has the potential to generate hate speech, which makes the detection of such content online challenging due to the complexity of the natural language constructs. Hate speech on social media not only degrades user experience but can also escalate to severe realworld consequences, including loss of life. Therefore, this highlights the necessity of implementing effective hate speech detection systems. There has been extensive research conducted in the field of machine learning, natural language and neural network techniques which are currently considered to be the most effective approaches for solving text classification related tasks. However, those approaches still need improvements to better detect hate speech online while protecting freedom of expression. This paper presents an empirical study on the application of capsule networks for classifying Twitter data into hate and non-hate categories. The capsule networks were originally developed for image classification tasks and have demonstrated strong performance in the respective domain. In recent times, the researchers have employed similar approach with respect to the text classification tasks. This study evaluates the robustness of the proposed model in comparison with well-known convolutional neural network and few machine learning techniques. The proposed model achieved 93 % accuracy than the experiments conducted on two widely used benchmarks. The proposed embedded capsule network model on the Twitter dataset showed exceptional results with highly competitive performance.

This paper presents a novel approach to semantically match “Resource Wanted” and “Resource Offering” classified ads within Sri Lanka's complex multilingual digital marketplace. We introduce a Siamese neural network architecture specifically designed to effectively process both textual content and categorical metadata across English and Sinhala languages. Our model leverages advanced multilingual transformer models to create semantically rich embeddings, with a LaBSE-based implementation achieving superior performance, reaching a Recall@1 of 0.5813 and a Recall@10 of 0.9151. Crucially, the integration of categorical features with text embeddings yielded the best results, demonstrating a ${1. 5 \%}$ improvement in Recall@1 over the text-only approach. Our methodology addresses the significant challenge of matching ads across linguistic bound-aries in a low-resource setting, providing a method that can significantly improve transaction efficiency in Sri Lanka's diverse digital marketplace.

Energy efficiency is becoming a crucial factor in addressing rising operational expenses, environmental sustainability, and energy supply constraints due to the increasing demand for computational resources and the rapid expansion of edge and cloud computing. Therefore, accurate workload prediction has been crucial for achieving energy efficiency in Edge and Cloud computing. This research evaluates the performance of three deep learning models namely, CNN-only, LSTM-only, and a hybrid CNN-LSTM for forecasting workload patterns using historical data. Evaluation metrics including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and $\mathbf{R}^{{2}}$ were used to assess model performance. The CNN-only model showed limited accuracy, reflecting its weakness in capturing temporal dependencies. In contrast, the LSTM-only model demonstrated near-perfect prediction accuracy due to its ability to model sequential data effectively. The CNN-LSTM hybrid model slightly outperformed the LSTM model, combining spatial feature extraction with temporal learning for enhanced results. These findings highlight the superiority of temporal or hybrid architectures in time-series workload prediction. Future work will focus on integrating attention mechanisms, deploying models in edge computing environments, and incorporating contextual features to improve model adaptability and real-world applicability in hybrid cloud-edge systems.

Establishing hydrogen as a clean and sustainable energy vector relies on advancements in solar-driven green hydrogen production technologies, such as photocatalytic water splitting (PWS). To make PWS commercially viable, it is essential to achieve a solar-to-hydrogen conversion efficiency of at least 10 %. Recent progress in computational materials science and machine learning has led to the development of sophisticated models such as graph neural networks and transformer-based architectures for identifying promising photocatalytic materials. However, their ability to accurately predict critical properties such as bandgap remains limited, primarily due to the lack of large, high-quality datasets computed using advanced density functional theory (DFT) functionals. While functionals like the modified Becke Johnson (mBJ) potential offer significantly improved accuracy, large scale datasets based on mBJ calculations are not yet widely available for training complex machine learning models. This study addresses these limitations by employing a novel transfer learning strategy. Models are first pretrained on large scale but less accurate Perdew-Burke-Ernzerhof (PBE) datasets and then fine-tuned using smaller, high fidelity mBJ datasets. This approach enhanced bandgap prediction accuracy by reducing the MAE by 10.22 % compared to a model trained conventionally on the mBJ dataset, thereby accelerates the discovery of efficient photocatalysts.

Social media platforms are now essential for communication, but they also lead to the quick spread of hate speech, which leaves people and society at serious risk. To lessen the influence of hate content, it is crucial to identify the top influential users who spread it. This research ranks the influence of hate users who propagate Twitter content. A novel method of Influence Score for measuring user influence was developed by integrating activity and engagement metrics, such as the number of hate tweets, retweets, likes, and replies. Users were divided into LOW, MEDIUM, and HIGH influence groups using k-means clustering, which allowed for a more focused examination of user behaviour. The findings showed that a smaller percentage of HIGH-influence users contribute to spreading hate content than users with LOW and MEDIUM influence. The validation process made clear how much better the suggested influence score was than followers, highlighting the value of using many metrics to find real influencers. The validation of the Influence Score was confirmed by comparison with crowdsourced rankings. With its scalable approach for detecting and classifying the influence of hate users, this study gives policymakers and social media administrators useful information. Platforms can significantly reduce the spread of hate content by targeting user clusters based on their influence levels and implementing actions based on the severity of each cluster. Future researchers can adopt this as a validation method for evaluating and improving their hate user ranking approaches.

As digital platforms expand globally, ensuring data sovereignty across regions has emerged as a critical concern, particularly in identity management systems. Existing identity architectures often centralize data storage and control, creating compliance risks in multi-jurisdictional deployments. This research proposes a novel architecture for multi-regional Identity and Access Management (IAM) systems to preserve data sovereignty while enabling unified customer identity experiences. The architecture ensures region-specific data storage, consent-aware data sharing, and regulatory compliance.

Cheating during online assessments is a major challenge faced by test administering institutions. Impersonation is the most common fraudulent act. Since impersonation can be resorted to at any stage of an online test, conventional access control at initial login is ineffective. Hence, the need for continuous authentication. Although technology is available to monitor candidates continuously while they take online tests, there are several constraints that limit their deployment. This paper presents a non-obtrusive, easily deployable, and effective continuous authentication system that could detect impersonation during online assessments. It uses a popular class of behavioural biometrics called ‘keystroke dynamics’ which are basically precision time measurements that can be made latently without the aid of any additional accessories. Authentication is performed using a novel parameter called the ‘angle of departure’ together with the ‘pause durations’, both computed from the biometrics captured during the online test and those captured during a mock test conducted as part of the candidate registration process. The underlying concept is presented in detail. Details of experiments conducted to establish proof of the concept are given. These include roleplays conducted for the acquisition of biometrics and simulations of genuine test taking as well as impersonation. The performance of the system is evaluated using both traditional metrics as well as recently established metrics that are considered as being more suitable for assessing continuous authentication systems. Analysis of the results indicates the effectiveness of the use of the angle of departure and pause durations for continuous authentication of candidates during online tests.

Manual interpretation of software deployment requests written in natural language remains a time-consuming and error-prone process in DevOps pipelines. Despite the growth of Infrastructure-as-Code (IaC) tools, translating ambiguous or incomplete user intents into executable deployment plans typically demands experienced human intervention. In this paper, we present a transformer-based intent classification framework using DistilBERT fine-tuned with Low-Rank Adaptation (LoRA) for lightweight, accurate interpretation of deployment-related queries. We introduce DeployIntentD, a custom dataset of 3,558 deployment-related queries, composed of real-world CI/CD logs and synthetic examples generated through Retrieval-Augmented Generation (RAG). Our model achieves 88.6% accuracy, 0.83 macro-F1 score, and sub-25ms inference latency on standard GPU hardware. The proposed system represents a step forward in bridging natural language understanding with DevOps automation.

Hate speech has grown to be a serious problem on social media, encouraging antisocial behaviour and platform abuse due to the ability to disseminate content to millions within seconds, often anonymously. This study contributes to the research community in two ways when addressing this problem. Firstly, a crowdsourcing platform is used to annotate YouTube data to assess the content’s severity level, which is used as the training dataset. Secondly, it uses XLM-RoBERTa, a transformer-based multilingual model designed for code-mixed and code-switched text that includes Sinhala, English, and Singlish, to automatically classify the severity level of hate speech. At the initial stage, YouTube text data was initially divided into hate and non-hate categories. Hate texts were then further divided into three severity levels: Bottom, Intermediate, and Top. The XLM-RoBERTa model was used to fine-tune a dataset of 15,000 YouTube titles and comments for multi-class classification. With an F1-score of 0.91 and an accuracy of 92%, the model outperformed deep learning techniques like LSTM and conventional models like Random Forest, Support Vector Machines, and Logistic Regression. This is regarded as the first experiment based on the “UN Strategy and Plan of Action on hate speech”, which offers a unique framework for automatically classifying hate speech according to severity. It gives governments, platform users, and policymakers useful resources to help them make sound decisions. Future researchers can utilize this technique to rank the influence of hate speech on hate users to counteract hate speech propagation at a more severe level.

Retrieval Augmented Generation (RAG) systems show promise for financial question answering, yet high accuracy on benchmarks such as FinanceBench ($\mathbf{1 9\%}$ baseline, 32% updated) remains challenging [1] [8]. This paper presents a systematic, multistage approach to significantly improve the performance of the RAG pipeline for financial QA. We first established a robust curated baseline using Gemini-2.0, Docling parser, Google's text-embedding-004, and a vector database, achieving an initial accuracy of $\mathbf{4 3\%}$. Subsequent architectural and component-wise optimizations were then iteratively implemented. Firstly, a metadata filtering strategy, which utilizes a fine-tuned NER model to extract company names and years from queries, improved accuracy to 72%, demonstrating that targeted retrieval can simulate the benefits of a single-store per-filing approach [1]. Secondly, a hybrid chucking technique, which preserves the structure of the document and utilizes tokenization sensitive refinements, further increased the accuracy to 80%. Third, the implementation of a Hybrid Search mechanism, combining dense and sparse retrieval methods, advanced performance to 84%. Finally, LLM-based query expansion, which transforms user queries into answer formats, yielded a final accuracy of 88%. This research demonstrates that a carefully designed RAG pipeline, incorporating intelligent metadata filtering, layoutaware chunking, advanced similarity search, and query semantics enhancement, substantially improves financial QA, significantly outperforming existing baselines.

The interplay between mood and eating has been the subject of extensive research within the fields of nutrition and behavioral science, indicating a strong connection between the two. Further, phone sensor data have been used to characterize both eating behavior and mood, independently, in the context of mobile food diaries and mobile health applications. However, limitations within the current body of literature include: i) the lack of investigation around the generalization of mood inference models trained with passive sensor data from a range of everyday life situations, to specific contexts such as eating, ii) no prior studies that use sensor data to study the intersection of mood and eating, and iii) the inadequate examination of model personalization techniques within limited label settings, as we commonly experience in mood inference. In this study, we sought to examine everyday eating behavior and mood using two datasets of college students in Mexico (N_mex = 84, 1843 mood-while-eating reports) and eight countries (N_mul = 678, 329K mood reports incl. 24K mood-while-eating reports), containing both passive smartphone sensing and self-report data. Our results indicate that generic mood inference models decline in performance in certain contexts, such as when eating. Additionally, we found that population-level (non-personalized) and hybrid (partially personalized) modeling techniques were inadequate for the commonly used three-class mood inference task (positive, neutral, negative). Furthermore, we found that user-level modeling was challenging for the majority of participants due to a lack of sufficient labels and data from the negative class. To address these limitations, we employed a novel community-based approach for personalization by building models with data from a set of similar users to a target user.

Intent classification is a foundational element in natural language processing, enabling conversational systems to accurately interpret user intent. In educational contexts, effective intent classification within Intelligent Tutoring Systems (ITS) can significantly enhance personalized student interactions. This paper presents the intent classification module for the Learner-Aware AI (LAAI) tutor, a dialogue-based ITS designed to recognize and respond to diverse student behaviors, such as valid answers, questions, expressions of boredom, and requests for clarification. We introduce LAAIIntentD, a custom data set specifically designed for this task, containing 1,244 labeled training records and 278 evaluation records. Leveraging this dataset, we fine-tuned a large language model (LLM) LAAI-intent-classifier using Low-Rank Adaptation (LoRA) techniques to create a lightweight yet powerful intent classifier. Our fine-tuned model achieves better overall Recall (0.86), Precision (0.85), and F1-Score (0.83) compared to GPT-based methods. GPT models with CoT and Few-Shot prompting improve Recall but sacrifice F1 scores. This highlights our model's efficiency in balancing accuracy and scalability for ITS applications.

The necessity for complex calculations in high-energy physics and large-scale data analysis has led to the development of computing grids, such as the ALICE computing grid at CERN. These grids outperform traditional supercomputers but present challenges in directly evaluating new features, as changes can disrupt production operations and require comprehensive assessments, entailing significant time investments across all components. This paper proposes a solution to this challenge by introducing a novel approach for emulating a computing grid within a local environment. This emulation, resembling a mini clone of the original computing grid, encompasses its essential components and functionalities. Local environments provide controlled settings for emulating grid components, enabling researchers to evaluate system features without impacting production environments. This investigation contributes to the evolving field of computing grids and distributed systems, offering insights into the emulation of a computing grid in a local environment for feature evaluation.

Enhanced Living Environment (ELE) applications utilize the expressive power of Artificial Intelligence (AI) to provide enhanced performance. However, despite the impressive performance, state-of-the-art AI techniques come at the cost of explainability; users cannot understand the rationale of the decisions made by such systems. This opaque nature causes a lack of trust leading to less adoption of ELE systems, especially in mission-critical domains such as healthcare. The AI community has proposed various eXplainable AI (XAI) techniques to rationalize AI decisions. However, the user perspective of XAI is not well explored. This chapter addresses user perception and requirements of XAI and attitude toward adopting explainable ELE systems. A user study with 326 participants revealed that most perceive XAI as essential and expect explanations to be easy to understand, faithful, and interactive. The respondents prefer concept-based explanations over feature attributions. Hence, we develop a novel approach to generate multimodal explanations consisting of linguistic and visual explanations to rationalize the decisions made by Human Activity Recognition (HAR) systems. Finally, we conduct a validation survey to evaluate the impact of introducing explanations into a HAR system. The results highlight that users' trust grows with explainability leading to higher adoption of ELE systems.

Text Segmentation involves dividing documents into thematically cohesive segments and it is essential for improving the effectiveness of document-level applications, such as summarization, question-answering, and information retrieval. In this paper, we propose a novel unsupervised approach for segmenting text content that combines advanced sentence embeddings generated by the Sentence-BERT model with document-level positional encoding and K-means clustering. By incorporating positional information, we enhance the embedding vectors to capture both the semantic relationships and the structural context of sentences within a document. This enriched representation aids in producing more contextually aware and coherent clusters. We evaluate the proposed method using the Choi dataset and compare the performance of clustering-based segmentation with and without positional encoding. Our experimental results show that the inclusion of positional encoding significantly improves segmentation quality, as indicated by the increment of the average V-measure from 0.53 to 0.73. This unsupervised approach provides a scalable and versatile solution for Text Segmentation, effectively bridging the gap between semantic coherence and document structure while eliminating the need for annotated training data.

Diverse communities of organisms inhabit environments ranging from the human digestive system to marine ecosystems, significantly impacting both human health and the environment. Metagenomic classification, a crucial concept in bioinformatics, can be performed at various taxonomic levels. In this research, we introduce a hybrid classification approach. By leveraging results obtained from reference database approaches, where classification is available up to the species level, we identify patterns to enhance the taxonomic depth of partially classified sequences. Our proposed solution integrates the composition and coverage information of gene sequences. Graph-based machine learning techniques show superior performance over traditional methods in hybrid sequence classification, as demonstrated by our experimental results.

The ALICE experiment at the CERN Large Hadron Collider relies on a massive, distributed Computing Grid for its data processing. The ALICE Computing Grid is built by combining a large number of individual computing sites distributed globally. These Grid sites are maintained by different institutions across the world and contribute thousands of worker nodes possessing different capabilities and configurations. Developing software for Grid operations that works on all nodes while harnessing the maximum capabilities offered by any given Grid site is challenging without advance knowledge of what capabilities each site offers. Site Sonar is an architecture-independent Grid infrastructure monitoring framework developed by the ALICE Grid team to monitor the infrastructure capabilities and configurations of worker nodes at sites across the ALICE Grid without the need to contact local site administrators. Site Sonar is a highly flexible and extensible framework that offers infrastructure metric collection without local agent installations at Grid sites. This paper introduces the Site Sonar Grid infrastructure monitoring framework and reports significant findings acquired about the ALICE Computing Grid using Site Sonar.

In the modern world, most software developers follow the microservice architecture for their software development. As a result, monolithic architecture-based software is also gradually transforming toward new microservice architecture. The main reason for using microservice architecture is to achieve the main software quality attributes such as scalability, maintainability, performance, portability, modifiability, and testability. Architects should consider product development and product deployment to achieve those quality attributes. Microservices must communicate with each other to produce reliable output, act as an independent service, and be deployed in distributed environments. The primary challenge in microservices architecture lies in software performance degradation, specifically in terms of response time and throughput, due to inter-service communication over the network. This research mainly focuses on the reference architecture for microservices to mitigate the challenges, which uses a request response-based steam TCP communication strategy for inter-service communication in both on-premise and cloud-native environments. People can achieve better performance and meet the main quality attributes with the suggested reference architecture.

ELMo was the first context based embedded generated neural model which mainly solved the problem of polesmy issue. Word embedding generated with ELMo was used with downstream tasks such as text classification to improve the classification performances. The architecture of ELMo consists of 3 layers such as convolution layer, and two LSTM layers. These layers can be either concatenate or used separately to generate different embedding representations. with the existing work, ELMo embedding layer performances was not properly evaluated for short text classification tasks. Moreover, it is hard to identify the best traditional classification algorithms with ELMo embeddings. Therefore, this study mainly focuses on identifying the best layer embedding weighting scheme as well as the best traditional machine learning algorithm which performs well with short text classification tasks. Seven short text type datasets were selected for the experiment. According to the experiment, concatenation of three layers was identified as the best layer embedding representation over the others. There was a around 2–4% performance improvement over the other layer combination. Moreover, the support vector machine algorithm showed the best classification performances among other Machine learning algorithms.