International Journal of Computer and Information Technology
Not a member yet
138 research outputs found
Sort by
Multi Moving Objects Detection in Video Using Pre-trained Deep Convolutional Neural Networks
Get PDFAbstract- Nowadays object tracking is a critical concern in the field of machine vision. With the advent of powerful computers, affordable cameras, and growing demand for automatic video analysis, researchers have shown significant interest in object tracking. Various methods have been proposed for tracking objects in machine vision, but a key challenge remains: ensuring the robustness of tracking algorithms across consecutive video frames. In recent years, deep neural networks have emerged as a promising approach for accurate position estimation. In this study, we propose an enhanced method that combines deep convolutional neural networks with established techniques like K-means clustering. Our approach addresses challenges such as object disappearances and severe displacements. The selection of deep neural networks is motivated by their compatibility with target identification in video sequences, and achieving a remarkably low error rate in tracking validates our claim
User Rating Prediction Method Based on Fine-tuning of Large Language Models
Get PDFOnline reviews in social networks reflect users\u27 preferences for specific attributes of products. Accurate predictions of user ratings based on these reviews can help businesses better understand genuine user feedback. The purpose of this study is to fine-tune large language models using online reviews and corresponding user rating data, generating a large model for predicting user ratings based on reviews. An attention mechanism is introduced to calculate attention weights for fine-grained review texts, reflecting the contribution of different text features to user rating prediction. By visualizing these weights, the process of calculating the predicted rating values can be explained. Experimental results show that the proposed interpretable user rating prediction method can effectively visualize the attention weights of important text features in the decision-making process of the large rating prediction model. Compared to the baseline model, the mean absolute error is reduced by 1.96, and the root mean square error is reduced by 1.73
Embedded Feature Selection Augmented Thyroid Disorder Prediction using MLP
Get PDFDue to its considerable fatality rate and increasing frequency, thyroid disorders pose a severe hazard to people\u27s health in the modern era. Thus, it has become a useful topic to predict thyroid disease early on using a few basic physical indications that are gathered from routine physical examinations. Being aware of these thyroid-related signs is crucial from a clinical standpoint in order to forecast outcomes and offer a solid foundation for additional diagnosis. However, manual analysis and prediction are difficult and tiring due to the vast volume of data. Our goal is to use a variety of bodily signs to swiftly and reliably predict thyroid disorders. This research presents a novel prediction model for thyroid disorders. We provide a deep neural network and embedded feature selection method-based algorithm for predicting thyroid disorders. Based on the LinearSVC algorithm, this embedded feature selection method selects a subset of characteristics that are strongly linked with thyroid condition by employing the L1 norm as a penalty item. We feed these features into the deep neural network that we constructed. To improve the performance of the predictor, gradient varnishing or explosion is avoided by initializing the network\u27s weight using the He initializer. The predictor is evaluated using a number of indicators like accuracy, recall, precision and F1-score. The results indicate that our model achieves 98.3%, 98.1%, 98.0% and 0.982 respectively, and that its average AUC score is 0.98, indicating that the approach we proposed is effective and trustworthy for predicting thyroid disorders
An Ensemble Predictive Model for Learner Attrition in Online Learning
Get PDFTo increase student retention and the success of online learning initiatives, it is critical to make very accurate predictions about learner attrition. In order to put early intervention strategies into place, universities must identify students who are likely to withdraw early. A number of variables, such as academic achievement, demographic traits, and engagement metrics, affect how accurately learner attrition is predicted. Effective prediction models will be developed by analysing these characteristics using machine learning techniques.
This study\u27s main goal is to create an ensemble-based machine learning model that predicts early learner attrition in Kenyan online learning environments by combining XGBoost, Neural Networks Decision Trees (DT), and Random Forests (RF). Learning Management Systems (LMS) secondary data collected from Kenya\u27s five universities will be used in the study. In order to provide a strong framework for the early detection of learners who are at risk, this study describes the technique for data preprocessing, feature selection, model training, and integration.
The research\u27s conclusions will help institutions and policymakers enhance online learning platforms, maximise student retention strategies, and tackle e-learning issues. The research intends to aid in the creation of a more effective and inclusive online learning system in Kenya by early detection of students who are at risk
A Lightweight 3D Convolutional Network for Hyperspectral–LiDAR Patch Classification
Get PDFWe propose a simple yet effective three-dimensional convolutional neural network (3D-CNN) for urban land-cover classification using co-registered hyperspectral imagery (HSI) and LiDAR data. The network treats the entire spectral-LiDAR stack as a three-dimensional volume and uses a series of 3×3×3 convolutions to capture both spectral and spatial context simultaneously. LiDAR elevation data is added as an extra channel in the input. During preprocessing, each hyperspectral band and LiDAR DSM are normalized to zero mean and unit variance. Training uses small local patches (P×P) centered on labeled pixels, with random flips and 90-degree rotations, called dihedral augmentation, applied across all channels. To address class imbalance, inverse-frequency class weighting and label smoothing are included in the cross-entropy loss. Evaluation on the Houston2013 dataset shows that the model achieves high accuracy, a single model reaches an Overall Accuracy (OA) of about 0.90 and an Average Accuracy (AA) of about 0.92 over five runs. An ensemble of five runs improves these results to OA ≈ 0.912, AA ≈ 0.928, and a kappa coefficient (κ) of approximately 0.904. Classes with distinctive spectral and spatial signatures, like water, synthetic grass, and tennis courts, reach nearly 100% recall. Meanwhile, classes with similar appearances, such as highway and road, show higher confusion, with highway recall around 46.9%. These results confirm that combining spectral and three-dimensional structural information significantly enhances accuracy in urban classification
An Android-Based Smart RSU Framework for Enhanced Urban Traffic Management
Get PDFThis paper addresses critical challenges in the deployment and effectiveness of traditional Roadside Units (RSUs) in traffic monitoring systems and proposes a novel, cost-effective approach using Android-based smart RSUs. Leveraging mobile phone architecture, YOLOv8 , the SAHI algorithm and chat gpt-4o, the system provides real-time traffic data collection, vehicle detection, and congestion analysis. This paper evaluates the performance of different cost tiers of mobile devices, discusses traditional traffic monitoring challenges, and identifies key gaps in current RSU technologies. The proposed system offers enhanced scalability, flexibility, and reduced cost, making it an ideal solution for urban traffic management
Ensemble Feature Selection for Network Intrusion Detection: Combining Information Gain and Random Forest with Recursive Feature Elimination
Get PDFNetwork intrusion detection systems (NIDS) are essential for protecting computer networks against cyberattacks. The selection of a nominal set of essential features that may adequately discriminate malicious traffic from normal traffic is indispensable while developing a NIDS. As such, a more reliable and accurate detection result may be realized when intrusion detection is carried out on a dataset based on an inclusive feature representation. This work presents the pre-processing and feature selection workflow as well as its results in the case of the CIC-IDS-2017 dataset, with a focus on two cyber-attacks, namely Denial-of-Service (DoS) and PortScan. The study applied an ensemble feature selection method based on information gain and Random Forest to filter out important features. Recursive Feature Elimination method was then applied to the reduced features to optimize the selected feature subset. The selected feature subset was experimented with using two classification algorithms, namely support vector machine and multi-layer perceptron. In the evaluation process, four widely used performance metrics were considered. The study results demonstrated the efficacy of the proposed ensemble approach to optimize the selected feature subset for detecting PortScan and DoS attacks in network traffic. Experimental results revealed that the support vector machine had a slight advantage in accuracy and could train more quickly. According to the study\u27s evaluation, the NIDS may be able to shorten processing times without sacrificing the ability to detect PortScan and DoS attacks accurately by choosing a narrow subset of informative features. This suggests the approach might be applicable to real-world NIDS scenarios involving these attacks. The study also provides encouraging perspectives on how ensemble feature selection utilizing MLP and SVM can enhance the effectiveness of NIDS. Building on these findings, more research can create NIDS solutions that are even more reliable and efficient for the dynamic field of cybersecurity
Aspect-Based Sentiment Analysis for Turkish Reviews Using Token and Sequential Classification Methods
Get PDFAspect-Based Sentiment Analysis (ABSA) aims to identify sentiments expressed toward specific aspects or attributes of entities in text. This study addresses the under-explored area of ABSA in the Turkish language by extracting aspect terms (targets) and their categories from customer reviews and determining the sentiment polarity for each aspect. Turkish, being a morphologically rich and structurally complex language, poses unique challenges that often hinder the direct application of methods developed for other languages. Hence, developing sentiment analysis approaches tailored to Turkish is of significant importance. We propose a two-stage pipeline: a token-level classification to recognize aspect terms and assign them to one of 12 predefined aspect categories, followed by a sequence-level (sentence-level) classification to predict sentiment (positive, negative, or neutral) for each identified aspect. We fine-tuned five transformer-based language models (BERT, ConvBERT, ELECTRA, DeBERTa, and DistilBERT) for aspect term and category extraction, and four models (BERT, ConvBERT, ELECTRA, DistilBERT) for sentiment classification. Experimental results on the SemEval-2016 Turkish ABSA Restaurant dataset show that the BERT model achieved the highest accuracy (92.20%) for aspect term and category identification, closely followed by ConvBERT (91.68%). For sentiment analysis, ConvBERT performed best with an accuracy of 86.91%, outperforming ELECTRA (85.34%), BERT (82.75%), and DistilBERT (77.48%). These findings demonstrate that pre-trained transformer models can effectively handle fine-grained sentiment analysis in Turkish, substantially improving on previous approaches. The proposed pipeline and comparative results provide a novel benchmark for Turkish ABSA, with potential applications in analyzing Turkish customer feedback to glean actionable insights
Automated Techniques for Detecting Healthcare Associated Infections: A Review
Get PDFAutomated detection of Healthcare-Associated Infections (HAIs) faces major obstacles due to unclear medical documentation, scarcity of well-annotated data, and multiple symptoms that overlap between HAIs. This review investigates recent advances in using classical machine learning, deep learning, transformers, and natural language processing (NLP) methods in detecting healthcare-associated infections. It examines empirical studies from 2019 to 2025, focusing on models\u27 performance based on various metrics, data issues, and ethical considerations. The study sought to assess and compare the performance of natural language processing (NLP) approaches of detecting Healthcare-Associated Infections (HAIs). Ethical and technical concerns such as data privacy and data imbalance, are critical barriers to implementation of NLP to detection of HAIs. The review underscores the promise of NLP to detection of HAIs while emphasizing the need for standardized metrics for evaluating HAI detection model and ethical frameworks of handling the datasets
The Development of the RSU U2 Net+ Architecture for Brain Tumor Segmentation in 3D Images
Get PDFSegmenting brain tumors in medical images plays a crucial role in diagnosis and monitoring of medical conditions. However, the segmentation process is still performed manually, consuming time and exhibiting variability among assessors. This research aims to develop the RSU U2-Net+ architecture for brain tumor multilabel segmentation in 3D images. The RSU U2-Net+ architecture consists of 9 interconnected blocks, employing broader connectivity in each block. The architecture is reinforced with the use of Residual U-blocks (RSU) to enhance image understanding across various scales without significantly increasing computational load. Testing on data reveals that the RSU U2-Net+ architecture performs well, as indicated by a dice coefficient score of 0.779, IoU of 0.6439, recall of 0.7541, and specificity of 0.9911. Evaluation is also conducted for each tumor label. Recall and specificity for edema are 0.8690 and 0.9851, for enhancing tumor are 0.7991 and 0.9956, and for non-enhancing tumor are 0.5942 and 0.9927. This research makes a significant contribution to the development of advanced medical image analysis technology. The achieved results have tangible benefits for medical practitioners and patients, with the potential to enhance the speed and consistency of brain tumor segmentation in 3D medical images