Clean and safe water is essential for human health and recognized as a fundamental right. However, the degradation of water quality worldwide, due to pollution and inadequate water management, poses serious risks to public health, making it crucial to evaluate water potability effectively. This study aims to classify water potability more accurately by leveraging machine learning models. The research utilized the Water Potability Dataset, which contains 3,276 records and 10 parameters, such as pH, hardness, and solids, all of which were analyzed to ensure a comprehensive assessment of water quality. Missing values in the dataset were addressed through imputation techniques, while class imbalance was resolved with data balancing methods. The dataset was divided into 80% for training and 20% for testing, ensuring reliable model validation. Decision Tree and Random Forest models were applied due to their efficiency in managing complex datasets and providing interpretable outputs. The Random Forest model outperformed the Decision Tree, achieving higher accuracy and demonstrating superior reliability in classifying water as potable or non-potable. Evaluation results indicate that these models effectively classify water as potable or non-potable, showcasing the potential of machine learning to improve water quality analysis. The findings of this study provide a reliable, data-driven framework that can support policymakers, water management authorities, and public health organizations in addressing water quality challenges. By facilitating timely and accurate assessment, this research contributes to safeguarding public health and promoting sustainable water management practices.
This study involved designing and implementing a compact educational blockchain system that facilitates understanding by eliminating much of the complexity of blockchain technology. The system, implemented in Python, ensures that the concept of blockchain integrity is easy to comprehend. Essential features include operation of consensus algorithms and use of block generation/addition to facilitate learning. The effectiveness of the system was evaluated via a survey of university students. The results revealed significant improvements in blockchain comprehension. These findings contribute to the development of foundational systems for blockchain education in academic settings.
Lung cancer remains to be a major cause of cancer-related mortality worldwide, that underscores the necessity of an early and precise diagnosis. This study investigates how feature selection can improve machine learning models for the diagnosis of lung cancer. Utilizing the Lung Cancer Dataset, we employed ranking-based feature selection methods, specifically Information Gain and Gini Decrease, for determining the most important features for model training. Three classifiers—Random Forest, Decision Tree, and Support Vector Machine (SVM)—were evaluated using measures including AUC, Classification Accuracy, F1 Score, Precision, Recall, and MCC. The results demonstrated notable improvements in model performance with feature selection, highlighting its importance in reducing computational complexity and improving diagnostic accuracy. The Random Forest model demonstrated the most favorable AUC performance of 0.980 and an accuracy of 94.4%, underscoring its robustness in predicting lung cancer outcomes. This study emphasizes the value of integrating feature selection into machine learning workflows to build efficient, interpretable, and clinically applicable models for lung cancer diagnosis.
Cricket is one of the most famous sports in the world that involves no competition alone, but a lively community of players, supporters, coaches, and technology development. There is a lack of automated systems offering useful, personalized feedback to enable the players to improve their techniques in trying to imitate their idols. This paper presents a data-agnostic framework that integrates data visualization, machine learning, and computer vision to allow users to analyze and emulate various stances of professional cricket players. Results obtained using a carefully curated dataset of classic cricket shots, including famous pull shots and cover drives, demonstrate how well the presented framework analyzes and contrasts batting strategies. It generates detailed results comprising a similarity comparison score and a two-dimensional visualization of posture differences, thus enabling users to further refine their techniques. Designed initially for cricket, the framework’s flexibility bridges the gap in emulating idols and allows for adaptation to other sports, demonstrating its broad applicability and innovative approach.
This paper presents a novel forest fire inventory dataset for Peninsular Malaysia, created using public remote sensing data spanning from 2001 to 2023. The dataset leverages the Google Earth Engine (GEE) framework established in our previous study, designed to address the lack of comprehensive, publicly accessible fire dataset for this region. It includes a total of 7349 columns (attributes) and 11,279 rows, consisting of 5650 fire points and 5629 non-fire points. The attributes encompass annual, seasonal, and monthly dynamic variables, as well as static, administrative, temporal variables, and class labels. As a result of this research, we have published and made the dataset publicly accessible at https://doi.org/10.5281/zenodo.11542164. This dataset is aspired to serve as a valuable resource for researchers and domain experts, facilitating further analysis and deeper understanding into forest fire behaviour in Peninsular Malaysia.
Threat intelligence is important for forming rapid, appropriate responses to increasing numbers of sophisticated cyberattacks. High-quality threat intelligence requires excellent threat information. Although intrusion detection systems that use deep learning are highly accurate, it is impossible to understand how they come to their decisions. Therefore, such systems cannot be used to created threat intelligence. Explainable artificial intelligence (XAI) is being intensively studied to clarify how deep learning decisions are made. However, standard XAI evaluation criteria are lacking, and comparative evaluations are impossible. This study aimed to establish XAI methods and evaluation criteria to determine whether it is possible to obtain high-quality threat intelligence. First, we carefully defined the role that a deep learning-based XAI intrusion detection system should play when creating high-quality threat intelligence. Then, we established XAI methods and evaluation criteria to examine the extent to which XAI fulfills this required role. The evaluation method was compared to current international standards. The results revealed that our criteria are generally applicable for assessing whether an XAI yields the information required to create high-quality threat intelligence.
This research investigates the integration of modern payment methods into a serverless water billing system leveraging Hyperledger Fabric. Following the successful implementation of real-time water consumption tracking on a blockchain network, this study aims to incorporate cryptocurrency payments through Coinbase Commerce and fiat payments via PayPal, thereby expanding payment options and enhancing system flexibility. The primary objective is to assess the feasibility, security, and efficiency of these payment systems within a decentralized billing framework. By offering cryptocurrency as an alternative for users favouring digital assets, while ensuring broader accessibility through PayPal, this research addresses the diverse needs of users. Additionally, it delves into the technical integration of these payment methods, focusing on transaction security, scalability, and user experience within the blockchain environment. Ultimately, this study aspires to develop a comprehensive model for modernizing water billing systems, incorporating secure, flexible, and efficient payment solutions that enhance customer satisfaction and operational efficiency.
The significant rise in social media usage has brought attention to the critical issue of fake news, which has become a prominent subject of academic and professional research in recent years. The detection of fake news written in Arabic presents distinct challenges due to the language’s structural complexity and instances of homography. These factors contribute to additional difficulties in analyzing the text and interpreting its meanings effectively. This research paper aims to improve the detection of fake news in Arabic. The authors employed various datasets sourced from multiple platforms. They implemented different machine learning and deep learning algorithms, such as Support Vector Machine (SVM) and Naive Bayes (NB), in addition to Long Short-Term Memory (LSTM) Networks and AraBERT. The results were analyzed, and the findings indicated promising outcomes, demonstrating enhanced accuracy in identifying fake news in Arabic.
This study aims to improve the predictive accuracy of sales and satisfaction analysis by addressing missing data using mean imputation and machine learning models. Numerical missing values were handled with mean imputation, while categorical missing values were excluded. The dataset was split into 80% for training and 20% for testing. Predictive models were built using Decision Tree, k-Nearest Neighbors (KNN), and Random Forest algorithms. Model performance was evaluated using metrics such as AUC, accuracy, F1-score, precision, recall, and MCC. Results demonstrate that the imputation process significantly enhanced model performance, with Random Forest achieving the highest AUC (0.999) and classification accuracy (0.982). This highlights the critical role of imputation in improving data quality and predictive reliability. Moreover, the study establishes Random Forest as a robust method for handling missing values and achieving superior predictive outcomes in similar datasets.
Finding out the consistency between two entities is important for grouping or learning similar behavior for classification and recommendations. Consistency means that an entity follows a certain similar pattern when related to other entities in the same group. Models are often classified according to categories, uses, complexity level, similar sets of classes, and other attributes. Heterogeneous models tend to have similar patterns and categories thus giving rise to an inconsistency. Identifying the similarity and consistency becomes crucial in Model-driven engineering where we often have to group UML/ECORE models for various purposes such as classification, Machine learning, and recommendation systems. In this paper, We have introduced the Siamese neural network to determine the similarity between the two models. This network carefully extracts the relevant features from the given models and generates the similarity index based on features and contextual information. The proposed approach has been validated using the Modelset dataset with an accuracy of 96.67% percent. The proposed approach sets a solid platform to evaluate model consistencies for different purposes.
This study developed a prototype for detection and identification of abaca plant pests and diseases. The creation of the prototype comprised both hardware and software development. In software development, the process was divided into two, namely model development and user interface (UI) development. During the model development, the pre-trained model of InceptionV3 CNN architecture was utilized to train the two classes of abaca diseases such as the abaca mosaic and abaca bunchy top virus (ABTV) and three classes of pests particularly the Brown Aphids, Slug Caterpillar, and Corm Weevil. The final model was integrated in the developed UI. The UI was developed using three different frameworks such as OpenCV, TensorFlow Lite and Flask. Moreover, for hardware component, the Raspberry Pi 4 was utilized as major component and as a microprocessor of the whole system. Upon testing the prototype, the performance achieved an accuracy rate of 94.6%. Its commendable performance shows its possibility to be deployed in a real-world detection and identification of abaca plants pest and diseases.
This study highlights the role of exhaled breath analysis as one promising, non-invasive technique used in the detection of cancer biomarkers, specifically focusing on Volatile Organic Compounds (VOCs) linked to lung and liver cancers. Three machine learning models were incorporated for model testing- Random Forest, Decision Tree, and SVM to identify the best classification models among these that distinguish the cancer profiles. The study achieved the highest accuracy of 95% for liver cancer and accuracy of rate 85% for lung cancer. Multivariate heatmaps and robust correlation matrices were plotted to highlight the unique VOC patterns which distinguish between benign, cancerous and healthy profiles. Advanced data imputation techniques were carefully evaluated, with zero imputation emerging as an effective method for optimizing data dimensionality. The VOCs associated with lung and liver carcinomas were condensed from 27 to 20 and 16 to 10, respectively, with compounds being shown as the most diagnostically significant. These findings support VOC profiling as a novel, innovative and patient-friendly approach for early cancer detection that could transform clinical oncology practices.
As global carbon emissions continue to rise, achieving carbon neutrality has become a primary objective worldwide. Predictive modeling for carbon emissions has emerged as a crucial strategy, with numerous methods developed to forecast and manage emissions effectively. Traditional models, which predominantly employ classical computing-based metaheuristic optimization techniques, often face limitations in reaching global optima, impacting prediction accuracy. This study introduces a hybrid quantum-classical approach to enhance prediction models, particularly through feature dimensionality reduction, which is essential for improving model performance. Quantum algorithms, especially quantum annealing (QA), have shown advantages in addressing complex optimization problems by utilizing principles such as superposition and tunneling, thus offering significant benefits over classical methods. The integration of QA with classical dimensionality reduction techniques, such as principal component analysis (PCA), represents a novel advancement in managing high-dimensional data within predictive models. By optimizing feature selection, our approach aims to increase both the accuracy and computational efficiency of carbon neutrality models. This paper proposes hybrid quantum-classical optimization opportunities that can serve as a foundation for future studies, ultimately leading to more robust and accurate predictive models.
The garment industry is a critical component of the global economy, and it has been a major driver of economic growth. The industry faces various challenges, including labour practices, one of which involves the estimation of the standard minute value (SMV). The SMV, representing the time required for a qualified operator to complete a task under standard conditions with appropriate allowances, is often estimated primarily based on engineers' experience. Different individuals may predict the SMV differently. Advancements in technology are expected to standardize SMV prediction and make the production processes more efficient. By knowing SMV accurately in advance, the garment production processes can be improved, thereby reducing cost of producing clothes. In this research, data are collected from a ready-made garment (RMG) industry, with the aim to apply machine learning (ML) based regression models to predict SMV outcomes without depending on industrial engineers. Among the regression models, linear regression (LR), decision tree regression (DTR), and random forest regression (RFR) are chosen for predicting SMV. For the model performance evaluation, mean square error (MSE) and squared correlation coefficient (R2) are calculated. The testing results showed that MSE values fall within 0.004 to 0.006 and R2 range from 0.77 to 0.86, indicating that ML-based regression models are quite accurate in predicting SMV. In addition to providing an efficient method for predicting SMV, this research helps in reducing manpower requirements, enhancing productivity, and minimizing losses.
Managing portfolios effectively in modern financial markets presents significant challenges due to rapidly changing dynamics and inherent risks. While multimodal machine learning models have gained traction across various disciplines, their application to portfolio risk analysis remains underexplored. This study introduces a novel methodology leveraging multimodal data sources to enhance portfolio risk management and forecasting. We employ the Temporal Fusion Transformer (TFT) model, a state-of-the-art deep learning framework, to forecast stock prices by integrating diverse data modalities, including historical stock prices, social media sentiment (Reddit discussions), financial news sentiment, and macroeconomic indicators. Risk metrics such as beta, volatility, Sharpe ratio, and Value at Risk (VaR) are computed and aggregated using a weighted approach. Distinct weighting schemes are applied to risk trends to capture varying market dynamics. The aggregated risk scores classify stocks into high, medium, or low-risk categories for each trend. By transforming multimodal financial data into actionable insights, this methodology enables traders to anticipate risks, optimize portfolios, and navigate volatile markets with greater confidence and efficiency.
Despite the availability of Artificial Intelligence applications across different domains, persistent usability challenges have hindered their effectiveness and user adoption. This systematic review investigates the role of interaction design in enhancing the usability and user experience of Artificial Intelligence applications. This analysis yielded three key findings. First, integration & interaction, learning, and communication are the primary Artificial Intelligence domains that leverage interaction design research to ensure user-centric and efficient applications. Second, Artificial Intelligence techniques such as hybridization, searching, and optimization are widely adopted to address multifaceted challenges. Finally, mixed methods have emerged as the most versatile user experience evaluation approach, integrating quantitative and qualitative techniques to provide comprehensive insights into Artificial Intelligence applications. These results offer a holistic framework for Artificial Intelligence practitioners and designers to develop intuitive and robust applications while addressing gaps in the field.
This study presents an in-depth comparative analysis of the performance of three key approaches in natural language processing (NLP) — transformers, recurrent neural networks (RNNs), and traditional machine learning methods — in the task of multiclass text classification within the legal domain in the Kazakh language. A specialized dataset for named entity recognition (NER), adapted to legal topics, was utilized for the analysis. The primary focus is on the classification of legal texts. Standard evaluation metrics, such as accuracy, recall, precision, and the area under the curve (AUC), were employed to provide an objective assessment of the models' efficiency in text classification tasks. Special attention was given to the processing of the Kazakh language, which belongs to the category of under-researched languages in computational linguistics. This highlights the necessity of developing specialized algorithms and adapting existing methods to work effectively with this language and its legal terminology. The conducted research not only expands the understanding of the capabilities of existing models for text processing in low-resource languages but also emphasizes the importance of further efforts to automate legal services. This can facilitate the development of more accessible and scalable tools for legal assistance.
Deep neural network training, particularly with Convolutional Neural Networks (CNNs), can be resource-intensive and time-consuming. Traditional early stopping strategies, which are primarily based on validation loss, may not always accurately capture the point at which a model has learnt the most relevant features, thereby resulting in overtraining or undertraining. This work presents Singular Vector Canonical Correlation Analysis (SVCCA) as a unique method for early stopping in CNN training. SVCCA provides a dynamic method for determining when a model has converged by examining the similarity of activations’ outcomes through SVCCA. The proposed method monitors the stability of the learnt representations and indicates when additional training no longer results in significant advances in model generalization. We show that SVCCA-based early stopping can result in more efficient training by reducing overfitting and preserving or even improving model performance. Experimental results on benchmark datasets show that the proposed strategy outperforms standard early stopping methods, providing a reliable option for optimizing training time in CNNs.
This systematic literature review examines the integration of Artificial Intelligence into STEM education, motivated by the need to understand how AI tools and methods are transforming educational landscapes. As AI technologies evolve rapidly, their potential to enhance teaching and learning within STEM disciplines warrants thorough investigation. This review delves into the applications of machine learning, deep learning, natural language processing, and reinforcement learning across various educational levels, drawing on 19 relevant studies identified through comprehensive database searches. The analysis reveals that AI significantly enriches teaching methods and student outcomes by offering innovative tools and methods for content generation, recognition, prediction, skill assistance, and evaluation, with a strong focus on enhancing mathematics education at the primary level. However, the limited number of studies included, and the absence of a meta-analysis highlight the need for ongoing research. Future studies are encouraged to expand the scope of literature searches and employ quantitative synthesis techniques to assess the impact of AI more robustly in STEM education. This review sets the stage for future explorations that could profoundly influence educational strategies and policies.
Health has become one of the top priorities in terms of research and impact metrics, with the target of finding relevant foods and strategies to improve public health problems. The utilization of plant agents such as vegetables, fruits, herbs, and nuts have been taken into attention for their various health benefits, including the inhibition of illnesses. Black garlic (Allium sativum L.) belongs to the group which has good medicinal properties and better taste than raw garlic. Black garlic gained popularity Ilocos Norte, Philippines and they have estimated annual production demand of eighty-five thousand hundred-gram packs. However, the manual classification of black garlic is hard and tedious to do out of the processing oven. This is due to lack of equipment in determining the aging quality. Hence, we created a device with Raspberry Pi 4 and camera module. The system acquires the images using artificial setup and applies trained ResNet50v2, VGGNet16, and EfficientNetB0 model that perform image classification. We trained the models using total of 1305 images of black garlic which was divided into 85/15 ratio training and validation subsets. The models were fine-tuned using less aggressive hyperparameters to adjust the already learned features of the model and introduce the new features of black garlic. The result achieved 98.33% accuracy for all the models upon testing the 60 samples of different class with errors in different samples. The EfficientNetB0 model is recommended for the use since it has the least model memory footprint at considerable same performance in terms of accuracy.
