Science, Technology, and Communication Journal

Systematic literature review using deep learning in plant genomic prediction

Rizki Darmawan — Mon, 01 Jun 2026 00:00:00 +0700

Genomic prediction (GP) is a critical approach in plant breeding used to forecast agronomic traits based on genetic marker data. In recent years, machine learning and deep learning methods have been increasingly applied in genomic prediction to address the limitations of traditional methods such as genomic best linear unbiased prediction (GBLUP). This systematic literature review aims to evaluate research trends over the past decade (2015–2025) regarding the application of deep learning in genomic prediction for crops, encompassing publication trends, publication sources, dataset types, key research topics, methods/algorithms used, implementation schemes, and commonly used evaluation metrics. The study follows the PRISMA protocol, with literature searches conducted on the primary databases Scopus and SpringerLink. From thousands of identified articles, 43 studies meeting the inclusion criteria were selected. The data revealed a significant increase in publications since 2018, peaking in 2024. The majority of articles were published in open-access journals, with notable contributions in Frontiers in Plant Science, Scientific Reports, and G3. The research covers various crop types, including wheat, maize, rice, soybean, sugarcane, and others, with diverse deep learning schemes such as convolutional neural networks (CNN), Ensemble Stacking, and Multi-Task Learning, often compared to conventional methods. The synthesis indicates that deep learning frequently enhances the accuracy of genomic prediction, particularly for complex multi-environment or multi-trait data. However, in some cases, classical linear/Bayesian models remain competitive. The application of deep learning in plant genomic prediction is rapidly advancing and demonstrates significant potential for improving crop selection performance.

Security audit of a pharmacy information system using blackbox testing and CIA triad: A case study

Rahmalia Syahputri, M Rivaldi Arwin Hadi Wijaya — Wed, 03 Jun 2026 00:00:00 +0700

Pharmacy information systems are essential for managing drug inventory, sales, financial reports, and user administration, yet they are exposed to security risks like data manipulation, account misuse, and information leakage. This study integrates Blackbox Testing and the CIA Triad (Confidentiality, Integrity, Availability) to audit a pharmacy application. Testing employed 19 security scenarios, supported by tools such as SQLmap, Burp Suite, OWASP ZAP, and Apache JMeter to detect vulnerabilities without accessing source code. Results show that the system meets availability requirements and provides audit logging for user activity monitoring. However, confidentiality and integrity weaknesses were identified: input validation allowed illogical data like negative stock, potential SQL Injection existed on the login page, and password encryption was insufficient. Strengthening input sanitization, adopting strong encryption, and enhancing authentication are necessary to close security gaps and improve system reliability.

QoS aware traffic shaping for TikTok live streaming over congested LAN using MikroTik queue tree

Diah Risqiwati, Wiryawan Ananta Pratama Panigoro, Hanugra Aulia Sidharta — Thu, 04 Jun 2026 00:00:00 +0700

Real time video streaming applications such as TikTok Live are highly sensitive to network instability, especially under bandwidth contention on shared access networks. This paper evaluates a Quality of Service (QoS)-aware traffic shaping scheme for TikTok live streaming in a congested local area network using a MikroTik router with a queue tree configuration. TikTok traffic is identified and assigned minimum bandwidth guarantees and highest priority, while non streaming traffic is treated as best effort. Network performance is assessed under two scenarios: baseline (without shaping) and experiment (with shaping), sing the TIPHON standard and four QoS parameters: throughput, delay, jitter, and packet loss. The experimental results show that although the average throughput only increases slightly from 70.67 Kbps to 80 Kbps and remains in the “Poor” category, the proposed scheme significantly improves temporal QoS metrics: average delay is reduced from 88.65 ms to 45.16 ms (“Very Good”), jitter decreases from 95.61 ms to 89.80 ms (“Good”), and packet loss drops from 7.66% to 3.78% (“Good”). These findings indicate that priority-based traffic shaping using a queue tree can effectively stabilize latency and data delivery for TikTok live streaming on bandwidth-limited networks without requiring capacity upgrades.

Robust cryptocurrency price forecasting using a Bayesian-optimized CNN–LSTM hybrid model

Tue, 09 Jun 2026 00:00:00 +0700

The rapid growth of cryptocurrency has caused the price movements of digital assets such as Bitcoin (BTC) and Ethereum (ETH) to become highly volatile and difficult to predict. This study aims to develop a cryptocurrency price prediction model using a hybrid convolutional neural network–long short-term memory (CNN-LSTM) architecture optimized with Bayesian Optimization. The data used in this study consisted of daily historical data for Bitcoin and Ethereum from January 1, 2018, to December 31, 2025, obtained from Yahoo Finance. The research stages included data preprocessing, normalization using Min-Max Scaling, sequence generation using the sliding window method (window sizes of 30, 60, and 90), CNN-LSTM model development, hyperparameter optimization using Bayesian Optimization (30, 50, and 100 trials), and evaluation using regression metrics including MSE, RMSE, MAE, MAPE, and R². The results showed that the hybrid CNN-LSTM model outperformed the standalone CNN and LSTM models, with RMSE reductions of 27% – 59% for BTC and 18% – 19% for ETH. For Bitcoin data, the best model was obtained using 30 trials with a window size of 30, achieving an RMSE of $2,588.33, MAE of $2,004.23, MAPE of 1.99%, and R² of 0.9468. Meanwhile, for Ethereum data, the best model was obtained using 50 trials with a window size of 60, achieving an RMSE of $138.56, MAE of $99.75, MAPE of 3.27%, and R² of 0.9737. These results indicate that the combination of CNN-LSTM and Bayesian Optimization is effective for predicting cryptocurrency prices with non-linear and volatile characteristics.

An IndoBERT-based framework for emotion classification in Indonesian song lyrics

Wed, 10 Jun 2026 00:00:00 +0700

Emotion classification in song lyrics represented a significant research area within natural language processing, yet studies targeting Indonesian-language lyrics remained scarce due to the limited availability of labeled datasets and the absence of domain-specific models. This study developed and evaluated an emotion classification model for Indonesian song lyrics using fine-tuned IndoBERT-base-p2, a transformer-based language model pre-trained on a large Indonesian corpus. A dataset of 1,025 labeled lyric entries was compiled from Kaggle, Genius, and KapanLagi, covering four emotion categories: joy, sadness, fear, and anger. Preprocessing encompassed duplicate removal, case folding, structural marker removal, and non-alphabetic character cleaning. Nine fine-tuning experiments were conducted by systematically varying learning rate and dropout rate, with early stopping applied based on validation loss. The optimal configuration employed a learning rate of 3 × 10^-5 and a dropout rate of 0.1, achieving 75.73% accuracy and 75.85% macro-averaged F1-score on the held-out test set. Joy and anger were classified most reliably, attaining F1-scores of 82.76% and 76.47% respectively, while sadness presented the greatest challenge, exhibiting the lowest precision of 64.10% alongside a recall of 80.65%, indicating a systematic tendency of the model to over-predict this class. These findings demonstrated that IndoBERT-base-p2, when fine-tuned with appropriate hyperparameter configuration, served as an effective approach for domain-specific emotion classification in Indonesian song lyrics.

Interpretative comparative analysis of LSTM and random forest for multi-label classification of English Qur’an translation

Nur Delifah, Nazruddin Safaat Harahap, Okfalisa Okfalisa, Elvia Budianita — Thu, 11 Jun 2026 00:00:00 +0700

The rapid growth of digital Qur'anic resources has created a need for automated systems capable of accurately categorizing verses by thematic content. The thematic complexity of Qur'anic text, in which a single verse may simultaneously convey multiple moral, spiritual, and social messages, presents a significant challenge for automated classification systems. This study conducts a comparative and explainable evaluation of long short-term memory (LSTM) and random forest (RF) for multi-label classification of English Qur'an translations across six thematic categories: arkanul Islam, iman, amal, human and community relations, akhlak, and history and story. To address severe class imbalance, synthetic minority over-sampling technique (SMOTE) was applied per label, expanding the training set from 4,489 to 19,658 samples. LSTM captured sequential contextual relationships through integer token embeddings, while RF relied on TF-IDF vector representations. Evaluated on 1,248 unseen test verses, RF achieved a higher macro F1-score (0.2748) compared to LSTM (0.2432), while LSTM retained marginally higher accuracy (79.61% vs. 79.55%). Per-label analysis revealed that both models performed best on lexically explicit labels such as arkanul Islam and iman, but consistently failed on abstract categories such as akhlak, where LSTM recorded near-zero recall of 0.61% and RF only 6.10%. This study contributes empirical evidence that TF-IDF-based SMOTE interpolation is more effective for minority-class augmentation than token-sequence interpolation, and demonstrates that macro F1-score is a more appropriate evaluation metric than accuracy for imbalanced multi-label religious text classification.

Workflow reproducibility and observability layer for Tomofast-x gravity inversion experiments

I Wayan Pio Pratama — Fri, 12 Jun 2026 00:00:00 +0700

Tomofast-x is an open-source parallel platform for gravity and magnetic inversion; however, reproducible execution, runtime observability, and workflow traceability are commonly managed outside the inversion software itself. This study presented a provenance-aware and telemetry-aware experimentation environment for reproducible Tomofast-x workflows and evaluated it using four publicly available reference scenarios archived on Zenodo. Each scenario was reproduced three times, resulting in 12 platform-managed runs executed through isolated workspaces with structured provenance and telemetry collection. The reproduced solutions showed close agreement with the reference outputs, with mean absolute root mean square error differences ranging from approximately 2.47 × 10^-11 to 4.72 × 10^-9. Runtime telemetry revealed substantial operational differences between scenarios. The uncompressed baseline required approximately 10.9 GB peak memory, whereas compressed scenarios required approximately 202 – 247 MB. Runtime decreased from approximately 436 s in the baseline case to approximately 300 – 340 s in compressed executions. Telemetry was successfully collected for all runs, including processor utilization, observed process-level RAM, runtime progression, and message passing interface (MPI) worker activity. Workflow robustness was further evaluated using injected failure cases involving corrupted parameter files, missing data grids, invalid execution paths, and simulated MPI failures. The results demonstrated that the proposed platform provided reproducible, provenance-aware, and telemetry-aware experimentation support for workstation-scale Tomofast-x workflows.

Adsorption of lead (II) ions using NaOH-activated matoa fruit shell (Pometia pinnata): Characterization and adsorption kinetics

Fadhil Maulana Harahap, T Abu Hanifah, Sofia Anita — Mon, 15 Jun 2026 00:00:00 +0700

This study focuses on the utilization of matoa fruit shell waste, which contains cellulose, as a potential biosorbent for binding heavy metals in solution. The study aims to examine the ability of matoa fruit shell powder (Pometia pinnata) as a biosorbent in removing lead (II) ions from solution and to analyze adsorption characteristics through kinetic studies. The research methods included biosorbent activation using NaOH at activation ratios of 1:1, 1:2, 1:3, 1:4, and 1:5 (w/v). The adsorption process was conducted with variations in parameters, including biosorbent dose, pH, and contact time. Characterization was performed using FTIR to determine functional groups, SEM-EDS to examine surface morphology and elemental composition, and ICP-OES to determine lead concentration in the solution. Kinetic analysis employed first-order pseudo-kinetic, second-order pseudo-kinetic, and intraparticle diffusion models. FTIR analysis results indicated the involvement of hydroxyl (-OH) and carboxyl (-COO^-) groups in the lead (II) ion binding process. The results of the study indicate that optimal adsorption conditions were achieved at a dose of 0.05 grams, a pH of 6, and a contact time of 60 minutes, with an adsorption efficiency of 90.76% and an adsorption capacity of 27.59 mg/g. The most suitable kinetic model was the pseudo-second-order model (R² = 0.9999), indicating a chemisorption mechanism.

Enhancing Indonesian hadith classification through multi-word embedding and support vector machine

Mila Hastati, Junadhi Junadhi, Susi Erlinda, Agustin Agustin — Tue, 16 Jun 2026 00:00:00 +0700

Hadith classification plays an important role in supporting the organization and retrieval of Islamic knowledge in digital environments. However, the increasing volume of digital hadith collections presents challenges for manual classification, making automated approaches increasingly necessary. This study proposes a hadith text classification framework based on support vector machine (SVM) and a Multi-Word Embedding approach. The dataset used in this study was obtained from the Kaggle hadith dataset repository and consists of 34,441 hadith records. The textual data were preprocessed through case folding, noise removal, stopword removal, and stemming before feature extraction. Three embedding strategies were evaluated, namely Word2Vec, FastText, and the proposed multi-word embedding, which combines Word2Vec and FastText representations through vector concatenation. The generated feature vectors were subsequently classified using SVM and evaluated using accuracy, precision, recall, and F1-score. Experimental results show that the proposed multi-word embedding approach achieved the best performance, obtaining an accuracy of 75.58%, precision of 75.68%, recall of 75.58%, and F1-score of 75.46%. These results outperform Word2Vec + SVM and FastText + SVM, demonstrating that the integration of contextual semantic and subword-level information produces richer feature representations and improves classification effectiveness. The findings indicate that multi-word embedding is a promising approach for automated hadith text classification and can contribute to the development of intelligent Islamic information systems.

The application of visual programming technology to classify mango leaves using a convolutional neural network

Silfia Andini, Teuku Radillah, Sumijan Sumijan — Fri, 19 Jun 2026 00:00:00 +0700

Leaf images are representations of leaves captured using digital cameras, which visually display the morphology and structure of the leaves. The classification of mango fruit can be identified from the type of leaves by using a convolutional neural network (CNN). The aim of this research is to obtain information on the type of mango fruit based on leaf type using image processing. The experimental setup involved the application of the test data set and the training data set to the classification of mango leaves, utilising iterations on both the training and testing data sets. This process involved the use of images exhibiting slight variations in shape, both in terms of image position and leaf type, to facilitate a comparison with images of the intended leaf type and leaves not included in the training classification. The experimental results yielded an accuracy level of 80.76%, validating the efficacy of the approach.

Antibacterial activity of a fraction of Senna alata L. leaves against Staphylococcus aureus: A comparative study of solvent types and concentrations

Eldya Mossfika, Dwi Saputra, M Saka Abeiasa, Azuxetullatif Azuxetullatif — Sat, 20 Jun 2026 00:00:00 +0700

Skin infections occur when microorganisms enter the bloodstream, spread to other organs, multiply, and cause severe illness. One of the common causative agents is Staphylococcus aureus. Thank to their phytochemical content, the Senna alata L. leaf have been used traditionally for antibacterial properties. Herein, we report the utilize of Senna alata L. leaf extract as an antibacterial agent against the Staphylococcus aureus grow up using a two-solvent fractionation method i.e. water and 96% alcohol at concentrations of 40%, 50%, 60%, and 70%. The phytochemical screening and disc diffusion tests were adopted to determine the phytochemical compounds and antibacterial efficacy of the fraction, respectively. Further, the statistical analysis including normality and homogeneity tests followed by a two-way ANOVA to evaluate the effects of solvent and concentration. The phytochemical screening detected the presence of alkaloids, flavonoids, tannins, and saponins in both fractions. The antibacterial efficacy results shows that the water and alcohol fractinations at a concentration of 70% produced the greatest antibacterial effect, with an inhibition zone diameter of 18 ± 0.18 mm and 18.90 ± 0.32 mm, respectively. A two-way ANOVA revealed a significant interaction (p = 0) between solvent type and concentration, indicating that antibacterial potency depends on both factors. These results confirm that the Senna alata L. leaf extract has a good potential for use as an antimicrobial agent.

Implementation of U-Net as EfficientNet encoder for brain tumour type classification

Rahma Aulia, Junadhi Junadhi, Lusiana Efrizoni, Rini Yanti — Mon, 22 Jun 2026 00:00:00 +0700

Brain tumor is one of the most dangerous diseases that requires fast and accurate diagnosis to support patient diagnose. The application of deep learning on magnetic resonance imaging (MRI) images has been widely used to assist automatic brain tumor classification. This study aims to implement a hybrid U-Net encoder-EfficientNet architecture for brain tumor classification using MRI images. In this study, the U-Net encoder was utilized to extract spatial features and generate an attention mask to highlight important regions before the classification process was performed by EfficientNet-B0. The dataset used was BRISC 2025, consisting of 6,000 MRI images divided into four classes: glioma, meningioma, pituitary, and no tumor. The experiments were conducted using three data splitting scenarios, namely 60:20:20, 70:15:15, and 80:10:10. The results showed that the proposed model achieved good classification performance across all testing scenarios. In the 60:20:20 scenario, the model achieved an accuracy of 82%, precision of 0.83, recall of 0.82, and F1-score of 0.81. In the 70:15:15 scenario, the model achieved an accuracy of 84%, precision of 0.85, recall of 0.84, and F1-score of 0.83. Meanwhile, the 80:10:10 scenario produced the best performance with an accuracy of 85%, precision of 0.86, recall of 0.85, and F1-score of 0.84. These results indicate that the use of the U-Net encoder was able to help the model focus on tumor regions, thereby improving the effectiveness of the classification process.

YOLOv11-based detection and classification of diseases in Siamese orange fruit using digital images

Rehan Khairuno, Anggi Hadi Wijaya — Wed, 24 Jun 2026 00:00:00 +0700

Diseases in Siamese orange fruit are one of the factors that can reduce the quality and yield of agricultural production. Manual disease identification requires considerable time and depends heavily on human observation skills; therefore, an automated system capable of detecting diseases quickly and accurately is needed. This study aims to implement the YOLOv11 model for detecting and classifying diseases in Siamese orange fruit based on digital images. The dataset used consisted of four classes, namely anthracnose, citrus canker, scab, and healthy, with a total of 627 images divided into training, validation, and testing datasets. The study utilized 20 variations of data augmentation, and DataV18 produced the best performance. The training process was conducted using the YOLOv11s architecture with 200 epochs and various data augmentation techniques. Based on the testing results, the model achieved a precision of 70.5%, recall of 61.2%, F1-score of 65.5%, mAP@0.5 of 61.8%, and mAP@0.5:0.95 of 44.8%. The results indicate that the YOLOv11 model has a fairly good capability in detecting diseases in Siamese orange fruit based on digital images and has the potential to be applied in the development of artificial intelligence-based plant disease detection systems.

Application of recursive feature elimination for sex classification of skull bones using random forest

Zam Afryan, Iwan Iskandar, Iis Afrianty, Benny Sukma Negara, Fadhilah Syafria — Wed, 24 Jun 2026 00:00:00 +0700

In forensic anthropology, sex estimation from the skull was a crucial initial step when visual identification of a body was not possible. Conventional methods relied on morphological assessment by experts, which was inherently subjective and dependent on the observer's experience. To address this limitation, this study implemented a computational approach using the random forest algorithm combined with the Recursive Feature Elimination feature-selection technique. The approach was evaluated using craniometric measurements from 2,524 individuals, comprising 1,368 males and 1,156 females, sourced from the Howell's craniometric dataset. The main challenge was the high dimensionality of the data, comprising 85 measurement features after non-biological attributes were removed. Using an excessive number of variables simultaneously introduced irrelevant information that lowered the model's ability to recognize true patterns, so the feature-selection technique was used to iteratively select the most informative measurements. The results showed that the model using all 82 features achieved an accuracy of 86.49 percent, while the optimized model using only 20 selected features achieved a higher accuracy of 86.85 percent. This indicated that by reducing the feature set by 75 percent, the model became lighter while remaining more accurate. The selection process further identified that cheekbone width and the height of the posterior ear protrusion were the most discriminative measurements between male and female crania, consistent with established biological evidence. In conclusion, the combination of random forest and recursive feature elimination produced an efficient and accurate sex-identification model, opening opportunities for its development as an objective forensic identification tool in the future.

Implementation of the mawaris fiqh hybrid chatbot based on retrieval-augmented generation and rule-based expert system

Fri, 26 Jun 2026 00:00:00 +0700

Islamic inheritance law (mawaris fiqh) regulates the distribution of inheritance based on the Quran, Sunnah, and ijma’. However, many people still have difficulty in understanding the concept of inheritance and performing accurate inheritance calculations due to the complexity of faraidh rules and limited sources of information about faraidh. This study aims to develop a hybrid-based mawaris chatbot that integrates retrieval-augmented generation (RAG) and rule-based expert system to support both conceptual question answering and deterministic inheritance calculations. This system is implemented using the Voyage-3-Large embedding model, Qdrant vector database, semantic caching, large language models (LLM) for contextual response generation using models from GPT-4o (main) and llama3.2:3b (fallback mode) as well as semantic cache using paraphrase-multilingual-MiniLM-L12-v2. The "Ask Concept" answering mode uses semantic search, confidence router, and RAG, while the "Calculate Inheritance" answering mode uses a rule-based expert system for heir identification, validation, faraidh calculation, and division result preparation. The system performance is evaluated for conceptual questions using BERTScore and weighted scoring model (WSM) for inheritance calculation questions. Experimental results show that the conceptual question-answering mode achieves a pass rate of 91.3% on questions in that domain. For inheritance calculation, the RAG-based approach achieves an average score of 44%, while the rule-based expert system achieves 100% in all evaluation categories. These findings indicate that the proposed hybrid architecture effectively combines the contextual reasoning capabilities of RAG with the deterministic accuracy of rule-based calculation, making it a reliable solution for mawaris consultation and inheritance distribution assistance.

Data-driven segmentation of sharia-based SMEs digital readiness: Comparing K-means and fuzzy C-means for strategic transformation planning

Ananda Vermiansyah, Okfalisa Okfalisa, Rahmad Abdillah, Surya Agustian — Fri, 26 Jun 2026 00:00:00 +0700

Quired to adopt digital technologies while preserving Islamic business principles such as transparency, fairness, trustworthiness, halal integrity, and ethical value creation. However, many sharia-based SMEs still lack clear diagnostic information regarding their digital readinesslevel, making it difficult for policymakers, business associations, and SMEs managers to design targeted digitaltransformation interventions. Prior studies on SMEs digitalisation have largely focused on technology adoption, digital marketing, or general readiness assessment, while limited attention has been given to data-driven segmentation models that can classify sharia-based SMEs into actionable readiness groups. Addressing this gap, this study compares K-means and Fuzzy C-means clustering to identify digital readiness patterns among sharia-based SMEs. The dataset consists of 314 SMES records collected through questionnaires and structured into two main perspectives, includes economic/business readiness and technological/digital readiness. The variables include business activity, transaction capability, management capability, market interaction, macro-environmental readiness, digital culture, digital education, financial resources, and technical infrastructure. Prior to clustering, the data were normalised to ensure comparability across indicators. K-means was used as a hard clustering benchmark because of its computational simplicity and ability to produce clear readiness groups, while Fuzzy C-means was applied as a soft clustering method because SMEs readiness boundaries are often overlapping and gradual rather than strictly separated. The clustering process was designed to generate three readiness categories viz., low, moderate, and high digital readiness. Model evaluation was conducted using silhouette index, Davies–Bouldin index, accuracy, F1-score, computational stability, and principal component analysis-based visualisation. The results show a trade-off between cluster separation quality and classification-oriented performance. Fuzzy C-means achieved a higher silhouette index of 0.1362 and a lower Davies–Bouldin index of 2.6126, indicating better internal cluster quality and stronger ability to represent overlapping readiness characteristics. In contrast, K-means produced higher accuracy of 0.6033 and F1-score of 0.3202, and more clearly formed three practical readiness categories. These findings suggest that Fuzzy C-means is more suitable for exploratory readiness profiling where SMEs may belong partially to more than one readiness stage, whereas K-means is more useful for managerial decision-making requiring crisp classification into low, moderate, and high readiness groups. This study contributes to SMEs digital transformation literature by demonstrating that sharia-based digital readiness should be analysed not only through aggregate scores, but also through segmentation models that reveal heterogeneous readiness patterns. Practically, the proposed comparative clustering framework provides a diagnostic basis for policymakers, Islamic business institutions, and SMEs development agencies to design differentiated digital capability-building programmes.

Nutri-score classification of snack products using word embedding and random forest

Onky Wanda Darmawan, Junadhi Junadhi, Lusiana Efrizoni, Nurjayadi Nurjayadi — Mon, 29 Jun 2026 00:00:00 +0700

The increasing consumption of packaged snack products has raised concerns regarding their nutritional quality and potential health impacts. Although nutritional information is commonly provided on food packaging, many consumers experience difficulties in interpreting ingredient descriptions and nutritional labels, making it challenging to identify whether a product is healthy or unhealthy. Therefore, an automated classification system is needed to assist consumers in understanding nutritional information more effectively. This study proposes a text-based classification framework for categorizing snack products into healthy and unhealthy classes using Natural Language Processing (NLP), word embedding techniques, and the Random Forest algorithm. The dataset was obtained from the Open Food Facts database and filtered to include snack products only. After preprocessing and class balancing, a total of 465 samples were used for model development and evaluation. The preprocessing stage consisted of case folding, tokenization, stopword removal, and stemming. Three word embedding techniques, namely Word2Vec, GloVe, and FastText, were employed to transform textual ingredient descriptions into numerical feature representations. Subsequently, Random Forest was utilized as the classification algorithm, and its performance was evaluated using Accuracy, Balanced Accuracy, Precision, Recall, F1-score, and Macro F1-score. The experimental results show that GloVe achieved the best performance among the evaluated embedding methods, obtaining an accuracy of 86.02%, balanced accuracy of 84.72%, precision of 85.98%, recall of 86.02%, F1-score of 85.91%, and macro F1-score of 85.19%. The findings indicate that GloVe provides a more effective semantic representation of food-related textual information compared to Word2Vec and FastText. Overall, the proposed framework demonstrates the potential of NLP-based approaches for automated nutritional assessment and healthy food classification.

Parametric optimization of an axial wind turbine blade for a hybrid renewable energy system integrating solar PV and micro-hydro

Tue, 30 Jun 2026 00:00:00 +0700

The growing demand for sustainable energy has driven the development of hybrid renewable energy systems that combine multiple sources to enhance reliability and efficiency. Solar photovoltaic (PV) and micro-hydro power are two widely adopted renewable sources, yet their performance is often limited by intermittency and seasonal variability. This research focuses on optimizing an axial wind turbine blade designed to harness wind energy from moving vehicles as a supplementary power source for a hybrid PV-micro-hydro system. The primary objective is to perform parametric optimization of the turbine blade to determine the optimal angle of attack that yields the highest lift-to-drag ratio, thereby maximizing aerodynamic efficiency. A computational fluid dynamics (CFD) approach using the FLUENT package was employed to simulate blade performance at various angles of attack (0°, 5°, 10°, 15°, and 20°), while maintaining constant wind speed, air pressure, and other parameters. The blade design features a diameter of 0.35 m, six blades, and a linear taper form, with testing conducted at a wind speed of 22 m/s, equivalent to the average speed of a moving truck on a highway. The simulation results demonstrate that an angle of attack of 10° produces the highest lift-to-drag ratio, indicating superior aerodynamic performance compared to other tested angles. Contours of velocity magnitude further confirm that the 10° angle yields the most favorable airflow distribution across the blade surface. The optimized blade design is now validated for integration as a supplementary wind energy component in a hybrid PV-micro-hydro system, contributing to increased overall energy output and improved system reliability. This research successfully achieves its parametric optimization goals, and the resulting blade design is ready for prototype assembly and further field testing.

A web-based decision support system for e-wallet selection using AHP-TOPSIS with integrated economic and technical values

Rahmat Zuhri Hafidz, Suwanto Sanjaya, Yelfi Vitriani, Fitra Kurnia, Febi Yanto — Tue, 30 Jun 2026 00:00:00 +0700

The rapid growth of financial technology in Indonesia has introduced a diverse range of digital wallet (e-wallet) options including OVO, GoPay, DANA, and ShopeePay. While this abundance of choice benefits consumers, it also creates decision-making challenges, particularly since most prior studies have neglected the balanced integration of economic and technical criteria in e-wallet evaluation. This study addresses that gap by developing a web-based decision support system for e-wallet selection using the AHP-TOPSIS method with integrated economic and technical criteria. Nine criteria were applied, comprising three economic criteria (cost structure, financial incentives, additional fees) and six technical criteria (data security, ease of use, merchant coverage, transaction speed, customer service, additional features). Primary data were collected from 85 active e-wallet users through a Likert scale 1 – 5 questionnaire. Results indicate that OVO ranked first with a TOPSIS preference score of 0.5835, followed by DANA (0.5802), GoPay (0.4654), and ShopeePay (0.3994). The developed system demonstrated the ability to produce objective, adaptive, and user-friendly recommendations to empower users with an interactive, data-driven decision support tool.

Predicting consumer loyalty from e-commerce reviews using emotion loyalty index with machine learning

Midrawati Hasibuan, Jeni Sukmal, Selamat Subagio — Tue, 30 Jun 2026 00:00:00 +0700

Customer reviews provide ratings and affective text, yet conventional sentiment classification and rating prediction do not offer a transparent multidimensional measure of consumer loyalty across e-commerce platforms. This study aimed to construct and evaluate a review-based emotion-loyalty index (ELI) for Shopee, Tokopedia, Lazada, Blibli, and Bukalapak. A quantitative computational design combined six formative components: normalized rating, lexicon sentiment, positive emotion, negative emotion, recommendation signals, and seller responses within a bounded 0-1 score. Reviews were preprocessed and represented using TF-IDF; corpus size was not reported in the supplied documents. Platform differences were examined descriptively, while Naive Bayes, Random Forest, Linear SVM, and Logistic Regression were evaluated on holdout data using accuracy and class-level F1. Blibli achieved the highest mean ELI of 0.363 and high-loyalty share of 53.8%, whereas Bukalapak recorded 0.254 and 24.9%, producing gaps of 0.109 and 28.9 percentage points. Tokopedia, Shopee, and Lazada obtained mean ELI values of 0.359, 0.336, and 0.317. Linear SVM reached 94.3% accuracy for emotion polarity, 93.6% for purchase intention, and 90.6% for lexicon sentiment, while Logistic Regression achieved 68.5% for loyalty tertiles. The proposed ELI contributes an interpretable framework that integrates evaluation, affect, advocacy, and seller engagement while separating platform comparison from leakage-controlled prediction. The framework can support platform diagnostics and targeted loyalty interventions across competitive digital marketplaces in Indonesia and comparable emerging economies, although future validation requires a timestamped corpus, complete inferential statistics, alternative weighting schemes, and longitudinal behavioral outcomes.

Identification of tenera and dura variety of oil palm fresh fruit bunches based on RGB color and fruit firmness

Tue, 30 Jun 2026 00:00:00 +0700

Sorting and grading oil palm fresh fruit bunches (FFBs) by variety in oil palm mills is destructive and inefficient, necessitating a more accurate, rapid, non-destructive approach. This preliminary study aims to develop a computer vision system to identify oil palm FFBs varieties (dura and tenera) using RGB intensities and fruit firmness levels. The study used 20 dura FFBs and 20 tenera FFBs, each with 10 ripe and 10 unripe FFBs, while fruit firmness was measured with a GY-3 needle-type penetrometer. Analysis of RGB intensities showed that ripe tenera had the highest values, while unripe dura had the lowest. In addition to RGB intensity analysis, this study also used principal component analysis (PCA) to visualize the separation patterns of varieties and ripeness levels based on RGB values. The PCA results showed that RGB intensity values clearly distinguished the dura and tenera groups in both ripe and unripe conditions. In terms of firmness, unripe fruits of both varieties had significantly higher firmness values than ripe fruits, with unripe dura showing the highest value of 12.5 kg/cm² and ripe tenera showing the lowest value of 7.23 kg/cm², indicating an inverse relationship between ripeness and fruit firmness. This study demonstrates that RGB intensities and fruit firmness levels can serve as potential parameters for distinguishing between the dura and tenera varieties in a computer-vision-based oil palm FFBs sorting system.

Application of ensemble methods on transformer sequence classification of BERT base uncased and RoBERTa-base models for hate speech detection

Reza Mahendra Sardi, Surya Agustian, Rahmad Abdillah, Febi Yanto — Tue, 30 Jun 2026 00:00:00 +0700

The rapid growth of social media platforms has brought a significant impact on the volume of digital interactions, which unfortunately is accompanied by a dramatic increase in the spread of hate speech and offensive language. Manual identification of negative content is highly inefficient and unscalable, thereby necessitating the development of state-of-the-art natural language processing (NLP) based automated detection systems. This study proposes the application of ensemble methods on Transformer architectures by combining two leading pre-trained language models, namely BERT (bidirectional encoder representations from transformers) base uncased and RoBERTa (robustly optimized BERT approach) base. The main focus of this research is to evaluate the performance of combining both models through a weighted average ensemble approach based on raw prediction probabilities (logits) with an even weighting ratio (50:50). Experiments were conducted using the public hate speech and offensive content identification (HASOC) 2021 dataset, covering two main scenarios: binary classification to distinguish NOT (normal) and HOF (hate/offensive) classes, and multi-class classification to categorize samples into HATE, OFFN (offensive), PRFN (profane), and NONE classes. To address the inherent challenge of significant class imbalance in the training data, this study implemented a custom class weighting function in the trainer module during the fine-tuning process. Empirical evaluation results demonstrate that the integration of the ensemble method effectively optimizes linguistic representation, suppresses prediction bias in minority classes, and improves performance stability. The ensemble model successfully achieved a macro F1-score of 0.8186 with 83.37% accuracy in the binary scenario, and a macro F1 score of 0.6570 with 68.93% accuracy in multi-class classification. This superior performance surpasses the capabilities of each baseline model individually, making it a robust hybrid architecture in tackling the variation of foul language in contemporary social media ecosystems.

Application of the smote and backpropagation neural network (BPNN) techniques in the classification of non- alcoholic fatty liver disease (NAFLD)

Tue, 30 Jun 2026 00:00:00 +0700

Non-alcoholic fatty liver disease (NAFLD) is a liver disorder with a high global prevalence and a significant mortality risk. However, clinical NAFLD datasets often exhibit severe class imbalance, causing machine learning models to become biased toward the majority class. This study aims to classify the mortality risk of NAFLD patients using the backpropagation neural network (BPNN) algorithm combined with the synthetic minority over-sampling technique (SMOTE). To ensure model validity, the follow-up time variable (futime) was excluded to prevent data leakage. The experiments were conducted by comparing different data split ratios (70:30, 80:20, and 90:10) as well as various hidden layer configurations and learning rates. The experimental results indicate that, without SMOTE, the model was trapped in the illusion of high accuracy (92%) while failing to detect mortality cases effectively (recall < 15%). In contrast, the application of SMOTE significantly improved the recall value, reaching 79.85% under the 80:20 data split scenario. These findings demonstrate that the integration of SMOTE and BPNN is highly effective in minimizing missed diagnoses (false negatives) in imbalanced medical datasets.

Apodization profiles of fiber Bragg grating spectra in temperature sensor applications

Tue, 30 Jun 2026 00:00:00 +0700

The presence of side lobes in the FBG reflection spectrum is a critical challenge that can reduce measurement accuracy in optical sensor applications, necessitating a systematic evaluation of apodization techniques as a mitigation strategy. This study investigated the comparative effects of three apodization profiles, namely uniform, Gaussian, and tanh hyperbolic, on the spectral characteristics and performance of an sensors-based temperature sensor with a grating length of 10 mm, grating period of 0.535 µm, modulation index of 0.0001, and operating wavelength of 1550 nm, over a temperature range of 20°C to 100°C. The results showed that all three profiles produced an identical thermal sensitivity of 13.725 pm/°C with a deviation of 3.23% from the theoretical value of 14.183 pm/°C. The Gaussian profile demonstrated the most superior side lobe suppression with an average side lobe level (SLL) of -45.51 dB, exceeding the minimum standard of -20 dB by more than 25 dB, far surpassing the uniform (-16.90 dB) and tanh (-18.17 dB) profiles. These findings establish the Gaussian apodization profile as the most optimal configuration, delivering the highest spectral quality without compromising thermal sensitivity.

Simulation Simulation and analysis of fiber Bragg grating for pulse strain monitoring using uniform and Gaussian apodization

Tue, 30 Jun 2026 00:00:00 +0700

Fiber Bragg Grating (FBG) is a passive optical sensor capable of detecting strain through a shift in the Bragg wavelength. This study designed and simulated an FBG-based sensor system for pulse-strain monitoring using two apodization profiles, uniform and Gaussian, at operating wavelengths of 1310 nm and 1550 nm. The sensor system was designed using OptiGrating 4.2.2.69 and OptiSystem 11.1.0.53 software with a single circuit configuration consisting of four parallel FBGs. The optical fiber parameters used included a core diameter of 2 µm, a cladding diameter of 8 µm, core refractive indices of 1.4600 and 1.4605, and a grating length of 10000 µm. The OptiGrating simulation results show that Gaussian apodization produces a cleaner reflection spectrum with lower sidelobes (around -60 dB) compared to uniform apodization, which exhibits higher and broader sidelobes (around -40 dB). In the OptiSystem analysis using an optical spectrum analyzer (OSA) at 1310 nm, Gaussian apodization yields an initial power of 24.920 dBm and a final power of 24.934 dBm, while uniform apodization yields an initial power of 25.000 dBm and a final power of 24.997 dBm. At 1550 nm, Gaussian produced an initial power of 24.725 dBm and a final power of 24.780 dBm, while uniform produced 24.998 dBm and 24.992 dBm, respectively. The strain value for both apodizations was calculated from the Bragg wavelength shift equation, yielding sensitivities of approximately 1.21 pm/µstrain at 1550 nm and 1.02 pm/µstrain at 1310 nm, consistent with theoretical FBG sensitivity. Optical power meter (OPM) measurements supported the OSA results with consistent power variations, and oscilloscope analysis showed stable signals for both apodizations, with uniform displaying slightly higher and more consistent amplitude. Uniform apodization excels in reflectance and amplitude stability, while Gaussian apodization excels in spectral quality with lower sidelobes and a higher strain-detection sensitivity that falls within the normal physiological pulse-strain range (10 – 50 µstrain). The choice of apodization therefore depends on the priority requirements of the strain-sensor application.