Articles

Access the latest knowledge in applied science, electrical engineering, computer science and information technology, education, and health.

Advanced Search
Filter Icon

Filters article

Years

FAQ Arrow
Single Year
Range Year
0
0

Source Title

FAQ Arrow

Authors

FAQ Arrow

31,291 Article Results

Time series forecasting: a comparative analysis of ARIMA, LSTM, and TFT models with missing data handling

10.11591/csit.v7i3.p291-303
Maryam Hosseini , Mohamad Forouzanfar
Time series forecasting plays a critical role in finance, healthcare, and energy applications, where accurate predictions support decision-making and operational efficiency. Traditional approaches such as autoregressive integrated moving average (ARIMA) perform well for linear patterns but often struggle with nonlinear and complex temporal dependencies found in real-world data. Although deep learning methods such as long short-term memory (LSTM) networks and Transformer-based models have shown promise, comprehensive evaluations across multiple domains and under missing-data conditions remain limited. This study compares ARIMA, LSTM, and temporal fusion transformer (TFT) models across three applications: stock price forecasting using SP 500 data, heart-rate prediction from electrocardiogram (ECG)-derived signals, and electricity demand forecasting using Pennsylvania–New Jersey–Maryland (PJM) power grid data. To evaluate robustness under realistic conditions, varying levels of missingness were introduced using missing completely at random (MCAR) and missing at random (MAR) mechanisms. Missing values were handled using forward fill, linear interpolation, and k-nearest neighbors (k-NN) imputation. Results show that TFT consistently achieved superior forecasting performance and demonstrated greater robustness to increasing missingness, while k-NN generally provided the most effective imputation performance across datasets.
Volume: 7
Issue: 3
Page: 291-303
Publish at: 2026-11-01

QuishingShield: on-device multi-modal detection of quick response phishing

10.11591/csit.v7i3.p271-290
Shabour Banda , Maronge Musara , Mainford Mutandavari
Quick response (QR) code phishing (quishing) attacks take advantage of user trust in QR physical and digital media, while currently available protection mechanisms only detect a single dimension signal and are unable to detect cross-modal deception. This paper introduces QuishingShield, an on-device quishing detection system based on multi-modal deep learning which preserves privacy on mobile platforms. Cross-modal attention fusion is used to combine visual poster features, optical character recognition (OCR) recognized surrounding text and recognized uniform resource locator (URL) structure, as well as network reputation signals in a system. A teacher model is trained on 205,488 real-world QR code poster samples from 45 countries and 23 languages, and the knowledge is distilled into a compact model for the student model, in order to be deployed in mobile applications. The student has an accuracy of 95.55% and a recall of 98.18% for a held-out test set, with an inference latency of 25.34 ms on mobile devices, all of which are at or below the deployment targets. Robustness of 95.00% when tested adversarially on four sets of attacks. The multi-modal fusion approach enhances performance by 5.17-11.07 percentage points over the unimodal baseline approaches (p0.001). QuishingShield, to our best knowledge, is the first validated multi-modal quishing detection system satisfying accuracy, speed, size and privacy requirement for mobile deployment.
Volume: 7
Issue: 3
Page: 271-290
Publish at: 2026-11-01

Hematological profiling of malaria-induced anemia using deep learning

10.11591/csit.v7i3.p304-313
Vanrose Panashe Nyamangodo , Wellington Makondo , Simbai Zindove
However, malaria-induced anemia (MIA) still persists to be one of the global health challenges with many cases of illness and fatalities mainly in pregnant women and children. Diagnosis of malaria and associated hematologic diseases such as anemia is traditionally carried out through examination of blood smear. However, such techniques require expertise, take long periods, and there are high chances of inter-observer variability. Here, an automatic system based on deep learning for detection of Plasmodium parasite and estimation of anemia is introduced. The proposed system uses convolutional neural network (CNN) branch for image analysis and multi-layer perceptron (MLP) branch for analyzing clinical data, thereby using their combination in multi-label classification. The advantage of such technique is the capability of the model to diagnose complex hematological signs and give probabilistic scores. This system was found to be accurate, precise and reliable.
Volume: 7
Issue: 3
Page: 304-313
Publish at: 2026-11-01

Early Escherichia coli prediction in broiler chickens

10.11591/csit.v7i3.p314-324
Nicole Chimwamafuku , Brian Mupini
Poultry farming remains an important contributor to global food security and commercial livestock production. However, infectious diseases such as Escherichia coli (E. coli) cause mortality, poor feed efficiency, reduced growth performance, and economic losses in broiler production systems. This study proposes a checkpoint-based multimodal transformer-convolutional neural networks (CNN) framework for early flock-level E. coli infection risk prediction using environmental, production, behavioural, and visual poultry data. Flock monitoring records collected from 2022 to 2025 were structured across six production checkpoints: day 3, day 7, day 14, day 21, day 28, and day 31. After long-format conversion, approximately 90,000 temporal observations were used for transformer modelling, with 72,000 records for training and 18,000 for testing. The CNN component evaluated 249 poultry images across healthy, low-risk, medium-risk, high-risk, and non-broiler classes. The transformer model achieved 99.96% accuracy, while the CNN model achieved 95.58% accuracy. The integrated dashboard generated flock risk scores, contributing factors, alerts, gradient-weighted class activation mapping (Grad-CAM) explanations, and veterinary advisory recommendations, demonstrating the potential of multimodal artificial intelligence (AI) for proactive poultry health monitoring.
Volume: 7
Issue: 3
Page: 314-324
Publish at: 2026-11-01

AI health assistant combining transformers and XGBoost for multilingual care

10.11591/csit.v7i3.p353-368
Shamiso Simango , Mainford Mutandavari
Limited healthcare access, shortages of healthcare professionals, and linguistic diversity continue to impede timely symptom assessment and healthcare delivery in low-resource settings such as Zimbabwe. Existing virtual health assistant (VHAs) are frequently cloud-dependent, English-centric, and lack interpretable decision-making, limiting their effectiveness in bandwidth-constrained and privacy-sensitive environments. This study proposes CIMAS HealthMate, a hybrid multilingual VHA that integrates transformer-based natural language processing (NLP) with an explainable extreme gradient boosting (XGBoost) decision model to provide accurate and transparent symptom triage. The framework employs the no language left behind (NLLB) model for offline English–Shona translation, bidirectional encoder representations from transformers (BERT)-based models for intent classification and medical entity recognition, and XGBoost for structured triage recommendation. The system was evaluated using a multilingual symptom corpus and an anonymized electronic health record-style dataset comprising approximately 23,000 patient records. Experimental results achieved translation accuracies of 76.5% for Shona-to-English and 82.2% for English-to-Shona, symptom extraction accuracy of 86.6%, and end-to-end triage accuracy of 93.3% with an F1-score of 93.3%. These findings demonstrate that the proposed hybrid architecture effectively combines multilingual language understanding, interpretable machine learning, and offline deployment to deliver reliable and privacy-preserving triage support. The proposed approach provides a scalable and practical solution for improving equitable digital healthcare services in multilingual, resource-constrained environments.
Volume: 7
Issue: 3
Page: 353-368
Publish at: 2026-11-01

Optimizing deep learning models for plant leaf disease classification using nature-inspired algorithms

10.11591/csit.v7i3.p369-376
Avinesh Culloo , Avinash Bhunjun , Geerish Suddul
Plant diseases greatly affect agricultural production, especially in developing countries, where prompt diagnosis can be quite challenging due to the limited availability of experts in real-time. Deep learning techniques for image analysis is gaining popularity and are increasingly considered an alternative to traditional manual inspection of plants. This research presents the evaluation of plant leaf disease detection system based on a convolutional neural network (CNN) optimized with different nature-inspired algorithms. The backbone model is based on the EfficientNet-B0 pretrained on ImageNet. Therefore, transfer learning is used to adapt the model to an updated PlantVillage dataset. Experiments have been conducted with multiple nature-inspired algorithms to improve generalisation and training efficiency of the prediction model. Different data preparation techniques have been carefully applied to the dataset, creating a unified approach to ensure consistency in the preprocessing pipeline for the training, validation, and testing phases. Our experiments indicate that application of the grey wolf optimizer (GWO) for tuning key hyperparameters of the model, including dropout, learning rates, and weight decay produced the best results, with and accuracy around 99.45%.
Volume: 7
Issue: 3
Page: 369-376
Publish at: 2026-11-01

A comparative review of modern large language model paradigms: GPT-4, BERT, Gemini, and DeepSeek

10.11591/csit.v7i3.p404-418
Kavish Sanghvi , Aparna S. Sharma , Surbhi Hooda
This review provides comparative analysis of GPT-4, BERT (bidirectional encoder representations from transformers), Gemini, and DeepSeek large language models (LLM), focusing architectures, training methodologies, and real-world applications. The primary research question is: How do these models differ in design, strengths, limitations, and potential areas for enhancement? By addressing this question, the study aims to provide insights into the trade-offs and future directions for optimizing LLM performance and deployment. The analysis reveals GPT-4 excels in natural language generation and complex reasoning, supporting up to 128K tokens with moderate latency and higher costs making it effective for conversational artificial intelligence (AI). BERT excels bidirectional contextual understanding with smaller computational overhead and broad open-source adoption, effective for text classification. Gemini demonstrates superior multimodal integration processing text, image, audio, and code with context lengths up to 1M tokens, enabling cross-domain adaptability. DeepSeek excels in specialized domains like finance and programming, is optimized for efficiency and supports extended context windows exceeding 200K tokens. However, all models face challenges related to computational cost, hallucinations, and ethical concerns, necessitating further improvements. Despite advancements, LLMs continue to grapple with issues such as data bias, model interpretability, and responsible AI deployment. Future research should focus on hybrid model approaches, domain-specific fine-tuning, and transparency to mitigate risks while maximizing the transformative potential of LLMs in real-world applications.
Volume: 7
Issue: 3
Page: 404-418
Publish at: 2026-11-01

Detection and translation of logic gate images into Boolean functions

10.11591/csit.v7i3.p346-352
Muhammad Shidqii Taqiyyuddin , Muhammad Subali
Artificial intelligence (AI) has become an essential tool in solving complex problems effectively and efficiently. In digital electronics, translating logic gate circuits into Boolean functions can be challenging, especially for more complex structures. This study presents the design of a detection and translation system for logic gate images into Boolean functions using the You Only Look Once (YOLOv5) object detection model. A dataset of 800 images was collected using a smartphone camera under varied lighting conditions and preprocessing to ensure robustness. The dataset was divided into 70% for training, 20% for validation, and 10% for testing. Training was conducted using YOLOv5s with batch size 32, 100 epochs, and pre-trained weights. The trained model achieved strong results with an overall mean average precision (mAP@0.5) of 0.922, precision of 0.98, and recall of 0.984. The confusion matrix confirmed accurate detection across all classes, with minimal misclassification. Furthermore, the translator system successfully converted recognized objects into correct Boolean expressions, with results validated in multiple test cases. This demonstrates that the proposed system can reliably automate circuit image-to-Boolean translation, bridging image recognition and symbolic computation.
Volume: 7
Issue: 3
Page: 346-352
Publish at: 2026-11-01

A hierarchical mixed-effects modeling framework with Weibull reliability characterization for spatiotemporal throughput variability in cellular networks

10.11591/csit.v7i3.p241-255
Victor Dela Gordon , Amevi Acakpovi
Reliable cellular throughput is essential for ensuring consistent user experience in modern mobile networks, yet it exhibits significant variability across spatial and operational conditions due to propagation effects, interference, and network congestion. This study proposes a hierarchical mixed-effects modeling framework integrated with Weibull-based reliability analysis to characterize spatiotemporal throughput variability in real-world operating conditions. The analysis is based on a large-scale dataset comprising over 77,000 field measurements collected across multiple university campus locations in Ghana, enabling cross-layer evaluation of network performance using key indicators, including reference signal received power (RSRP), reference signal received quality (RSRQ), and round-trip time (RTT). Analysis results indicate that all modeled predictors significantly influence throughput performance, with signal quality emerging as the dominant factor, alongside notable spatial heterogeneity. The model explains 18.9% of variability using fixed effects and 45.1% when spatial effects are included. Reliability analysis indicates that the probability of achieving 5 Mbps and 10 Mbps is 39.7% and 22.5%, respectively. These findings demonstrate that the proposed framework effectively captures throughput variability, spatial heterogeneity, and probabilistic service reliability in operational cellular environments, providing a practical analytical framework for reliability-aware cellular network optimization and performance evaluation.
Volume: 7
Issue: 3
Page: 241-255
Publish at: 2026-11-01

A literature review of recurrent neural network approaches to malaria outbreak prediction in Sub-Saharan Africa

10.11591/csit.v7i3.p377-393
Sophia Tembure , Wellington S. Manjoro
Malaria remains a significant public health issue in Sub-Saharan Africa. Predictive modelling is increasingly viewed as a way to move from counting cases reactively to preparing proactively for outbreaks. This literature review examines peer-reviewed research published from 1997 to 2025. It covers recurrent neural networks (RNNs) methods, attention mechanisms, multi-source data integration, and health-system informatics relevant to malaria prediction in low- and middle-income countries. A thorough search through Google Scholar, IEEE Xplore, ScienceDirect, PubMed, and SpringerLink identified 187 candidate papers. After reviewing titles and abstracts, 42 papers met the inclusion criteria. The review organizes the selected studies into six main categories: climate-disease ecology, statistical forecasting, machine learning (ML) baselines, RNNs, attention mechanisms, and health-system data integration. A comparative matrix highlights the similarities and differences in methods across twenty representative studies. The review points out four persistent gaps: the lack of attention-augmented RNNs for malaria forecasting in Southern Africa, limited integration of health-facility infrastructure features with climate predictors, inadequate handling of missing data in African satellite-derived climate series, and the absence of operational dashboards that make model outputs usable for district health officers without statistical training. Future research should focus on cross-country transferability studies, using graph neural networks to capture spatial spillover, and real-time integration with national surveillance systems like district health information software 2 (DHIS2).
Volume: 7
Issue: 3
Page: 377-393
Publish at: 2026-11-01

Risk-integrated contractor allocation in Zimbabwe’s timber value chain

10.11591/csit.v7i3.p337-345
Tavengwa Norman , Brian Mupini
In Zimbabwe, the commercial forestry industry is reliant on a large proportion of outsourced harvest and milling contractors whereas existing enterprise resource planning (ERP) systems record completed transactions instead of predicting contractor failure before assigning activities. We introduce the dynamic resource allocation framework (DRAF), a risk-integrated decision-support framework that supplements the probability of contractor failure derived from XGBoost to a non-dominated sorting genetic algorithm II (NSGA-II) multi-objective optimizer. It divides contractor-block-mill combinations, minimizes cost and expected delay, and maximizes risk-adjusted timber recovery and operational reliability. We created the solution on an 828,789-record virtual ERP dataset and calibrated it to Manicaland forestry conditions and tested with statistical, heuristic and risk-free optimization baselines. Extreme gradient boosting (XGBoost) obtained a holdout receiver operating characteristic area under the curve (ROC-AUC) of 0.965, recall of 0.999, and F1 of 0.867, showing an improvement of 0.365 over logistic regression. The optimizer developed 64 complete Pareto solutions to balanced and high-recovery scenarios and uncovered a constraint-feasibility boundary for a more conservative low-risk scenario. The satisfaction score of the 30-practitioner stakeholder assessment was 4.19 out of 5.0. The results demonstrate that embedding predictive risk into an optimization objective can optimize forestry allocation decisions and suggest that some real ERP validation is required before such measures will be broadly implemented.
Volume: 7
Issue: 3
Page: 337-345
Publish at: 2026-11-01

Hybrid geostatistical and machine learning for gold grade estimation

10.11591/csit.v7i3.p394-403
Tanyaradzwa Miriam Mtetwa , Monika Gondo
The evaluation of ore grade is a basic part of the process of mine planning, production scheduling and resource evaluation. The nugget effect and irregular mineralisation that occurs in some greenstone belt deposits on the Zimbabwean Archean causes problems in estimating accurately. The traditional geostatistical models like Ordinary Kriging (OK) produce smoothed estimates which have systematic underestimation in high-grade areas and overestimation in low-grade areas, thus affecting the deposit selectivity. In addition, machine learning (ML) techniques can capture the more complex nonlinear grade relationships, but they are unable to model spatial continuity and can be unrealistic in predicting grades. The hybrid OK-Gradient Boosting (GB) model was designed and evaluated with the drill-hole data of a gold mine in Zimbabwe. The spatial baseline was produced using OK and GB was trained to correct the systematic residual errors. The hybrid model was found to have the maximum overall explanatory power (R²=0.8795), competitive prediction accuracy and increased spatial realism. A spatial prediction map and extraction priority zone classification was created to aid operational mine planning.
Volume: 7
Issue: 3
Page: 394-403
Publish at: 2026-11-01

Convolutional neural network and long short-term memory forecasting and variational autoencoder anomaly detection in 4G cellular networks

10.11591/csit.v7i3.p256-270
Ruvarashe C. Hove , Eng Mainford Mutandavari
Spectrum monitoring in cellular networks with limited resources remains predominantly reactive because capacity planning and failure detection occur after network congestion, while spectrum monitoring data are rarely available for reproducible research. Existing studies address short-term traffic forecasting and anomaly detection as separate tasks and largely depend on densely sampled operator telemetry data that are inaccessible to academia and regulators in emerging markets. This study integrates both tasks into a unified deep-learning pipeline for a 4G cellular environment. The Zimbabwe spectrum dataset contains 315,247 hourly measurements collected from 13 cell sites, three operators, and five frequency bands, annotated with four operator-defined anomaly classes representing 2.03% of all measurements. A hybrid one-dimensional (1-D) convolutional neural network-long short-term memory (CNN–LSTM) model uses a 72-hour traffic window to forecast the next six hours of aggregate network traffic, while a variational autoencoder (VAE) trained exclusively on normal records detects anomalies when reconstruction error exceeds the 99th-percentile validation threshold. On the held-out test set, the model achieved a mean absolute error (MAE) of 158.97 GB, root mean square error (RMSE) of 230.41 GB, and mean absolute percentage error (MAPE) of 52.93%, outperforming a seasonal-naive baseline (MAE=216.43 GB, MAPE=66.77%, p0.001, Diebold-Mariano test). The study contributes a publicly available localized synthetic dataset, a reproducible end-to-end forecasting and anomaly detection baseline, and a per-class anomaly analysis.
Volume: 7
Issue: 3
Page: 256-270
Publish at: 2026-11-01

A network-based security information system for safeguarding computer-based test platforms in organizational environments

10.11591/csit.v7i3.p325-336
Ajani Dele , Owolabi Abdulhakim Adewale , Inaya Adesuwa
Computer-based testing (CBT) platforms have transformed education and certification by enabling scalable, efficient, and accessible examinations. However, these systems face significant cybersecurity risks, including unauthorized access, denial-of-service (DoS) attacks, and digital cheating, which threaten fairness and reliability. This study proposes a network-based security information system (NBSIS) designed specifically for CBT environments. The framework integrates layered defense, including pfSense firewalls (FW), Snort intrusion detection, Splunk security information and event management (SIEM), and artificial intelligence (AI)-powered analytics, into a unified architecture. A human-centered dashboard ensures usability for non-technical exam administrators, providing real-time alerts and intuitive controls. Validation through simulated attack scenarios demonstrated strong resilience, with high detection accuracy, reduced false positives, and rapid response times. Comparative analysis against intrusion detection system (IDS)-only and SIEM-only systems confirmed superior performance. The findings highlight NBSIS as a robust, scalable, and adaptive solution that safeguards exam integrity while remaining practical for diverse organizational contexts. This research contributes to computer science by advancing secure architecture, applying AI-driven anomaly detection, and integrating human-computer interaction principles into cybersecurity for education.
Volume: 7
Issue: 3
Page: 325-336
Publish at: 2026-11-01

Design science research in developing a religious chatbot based on Bulugh al-Maram

10.11591/ijece.v16i5.pp2782-2794
Aris Tjahyanto , Irmasari Hafidz , Faizal Johan Atletiko
Chatbots have recently gained significant popularity. For instance, ChatGPT has become a preferred tool for many individuals seeking instant answers without relying on human responses. This immediacy sets chatbots apart from books, which require users to search for information manually. This time-consuming process does not align with millennials' preference for convenience and efficiency. Studying hadith independently using the Bulugh al-Maram book demands considerable time and effort. The limited use of natural language processing technologies in religious chatbots restricts their ability to handle complex inquiries effectively. A chatbot capable of answering hadith-related questions could greatly assist the public in studying hadith texts by providing direct responses without extensive searching. This chatbot was designed for web browsers, utilizing deep learning as its core technology. This research led to the development of a chatbot prototype for learning hadith from Bulugh al-Maram. Built using the design science research (DSR) methodology, the prototype achieves an intent recognition rate (IRR) of 86.82%. However, its capabilities are below the BERT model, demonstrating a strong ability to accurately interpret user questions and statements.
Volume: 16
Issue: 5
Page: 2782-2794
Publish at: 2026-10-01
1 2 3
...
2086 2087
Show 1 of 2087

Discover Our Library

Embark on a journey through our expansive collection of articles and let curiosity lead your path to innovation.

Explore Now
Library 3D Ilustration