41
Papers Presented
15+
Countries
150+
Authors
Artificial Intelligence for Sustainable Waste Management: A Literature Review
Authors: Syed Tasleem Ahmed (Alva's Institute of Engineering and Technology)*; Kiran Raj K M (Alvas Institute of Engineering and Technology); Pooja H (Alvas Institute of Engineering and Technology); Priyanka N (Alvas Institute of Engineering and Technology); Rohith Vilas Reddy (Alvas Institute of Engineering and Technology)
Emails: ahmedtasleem296699@gmail.com; kiranraj@aiet.org.in; harishpooja260@gmail.com; priyanka3516@gmail.com; rohithvilas4@gmail.com
Abstract:
This study cites developments in artificial intelligence (AI) applications in solid waste management (SWM); it employs major approaches to neural networks, the Internet of Things, and evolutionary algorithms. Integration with Industry 4.0 technologies has led to efficient, scalable, and sustainable waste management solutions, such as the Internet of Things, robots, and blockchain.Artificial Intelligence (AI) is revolutionizing waste management by leveraging real-time monitoring, predictive analytics, and waste-to-energy technologies. AI-based solutions optimize waste collection, minimize landfill reliance, and ensure sustainable waste processing. By combining AI with smart sensors and automation, cities can improve waste segregation, reduce human intervention, and eliminate operational inefficiencies. AI-based forecasting models also allow municipalities to predict waste generation patterns, enabling proactive management strategies that minimize environmental footprint and maximize urban cleanliness. Moreover, a discussion has been made of IoT and the creation of smart garbage containers which will optimize the waste collection route. The study further highlights evolutionary backbone approaches for addressing complex SWM problems, such as genetic algorithms and particle swarm optimization. It also attests to the need to take note of performance criteria, such as segregation efficiency and recycling rates, in managing and improving existing waste management systems. With an AI, cities can move to a circular economy while protecting the environment and increasing the well-being of society.
SymptoTech: Integrative AI System For Symptom Analysis And Customized Health Solutions
Authors: Utkarsha Chore (SRM Institute of Science and Technology,KTR)*; Ehteshamul Haque (SRM Institute of Science and Technology,KTR)
Emails: up3406@srmist.edu.in; eh8302@srmist.edu.in
Abstract:
SymptoTech is a comprehensive AI-based healthcare solution designed to offer accurate, personalized diagnostic insights. Specifically developed to address various health conditions with an emphasis on dermatology, SymptoTech integrates image analysis with symptom-based diagnostics to enhance healthcare accessibility. The system employs Convolutional Neural Networks (CNNs) for identifying skin conditions from user-provided images while incorporating Natural Language Processing (NLP) for additional symptom analysis, leading to improved diagnostic precision. Furthermore, SymptoTech features a dietary recommendation engine that provides users with health-specific nutritional guidance. By seamlessly combining data acquisition, AI-driven diagnostics, and personalized health insights, SymptoTech aims to bridge the gap between technology and efficient medical decision-making, ensuring a more accessible and precise healthcare experience.
Syntactic Graph Co-attention Network for Automatic Short Answer Grading
Authors: Onkar Sabnis (IIT Kharagpur)*
Emails: onkar.sabnis.29@gmail.com
Abstract:
In this work, we addressed the problem of Automatic Short Answer Grading (ASAG). The task involves assigning a grade to a student's answer by comparing it against a model answer for a given question. Previous works in this domain mostly used rule-based and machine-learning methods to tackle the problem, wherein the creation of handcrafted features and the use of neural networks have been the most common practice. Different variations of syntactic and semantic similarity between a student and model answer pair have been used as features in earlier works. We hypothesize that the extent of alignment between the graph representations of a student and model answer is a good indicator of their relative similarity. In this direction, we propose an end-to-end ASAG system that models the alignment as co-attention between the nodes in the dependency graphs corresponding to an answer pair. We leveraged the representational power of BERT and Graph Convolutional Network (GCN) along with a co-attention mechanism to capture the intrinsic similarities between student answers and reference answers. Our proposed method surpasses most of the existing state-of-the-art results on the SemEval-2013 SciEntsBank and BEETLE datasets.
Fine-tuning BERT with Kolmogorov-Arnold Network for Hierarchical Text Classification
Authors: Djelloul BOUCHIHA (University Centre of Na ma)*; Sofiane BOUKLI-HACENE (University of Sidi Belabbes); Abdelghani BOUZIANE (University Centre of Na ma); Abdelkrim BERRABAH (University Centre of Na ma); Mohammed Amine BOUCETTA (University Centre of Na ma); Benamar HAMZAOUI (University Centre of Na ma)
Emails: bou_dje@yahoo.fr; boukli@univ-sba.dz; bouziane@cuniv-naama.dz; berrabah.abdelkrim@cuniv-naama.dz; Boucetta.mohammedamine@cuniv-naama.dz; hamzaoui@cuniv-naama.dz
Abstract:
Hierarchical text classification (HTC) remains a challenging task in natural language processing (NLP) due to the complexity of multi-level label structures. Traditional deep learning models, such as BERT-BiLSTM, often struggle to effectively capture hierarchical dependencies. In this paper, we propose a novel approach that fine-tunes BERT with Kolmogorov-Arnold Network (KAN) to improve HTC performance. BERT is leveraged to generate contextual text embeddings, while KAN, a neural network variant with learnable activation functions on edges instead of fixed activation functions on neurons, enhances classification by capturing hierarchical relationships more effectively. We evaluate the BERT-KAN model on a large-scale dataset of Amazon product reviews, structured into three hierarchical levels. Our results demonstrate that BERT-KAN outperforms BERT-BiLSTM across multiple evaluation metrics, particularly in Category Accuracy and Hierarchical F1 Score, confirming its superior ability to model hierarchical structures.
Deep Learning for Arabic Question Classification: Leveraging AraELECTRA and CNNs
Authors: Soumia KHEDIMI (University Center of Naama)*; Abdelghani Bouziane (Ctr Univ Naama, Dept. Mathematics and Computer Science, EEDIS Lab., UDL-SBA, Algeria); Djelloul BOUCHIHA (Ctr Univ Naama, Dept. Mathematics and Computer Science, EEDIS Lab., UDL-SBA, Algeria)
Emails: khedimi@cuniv-naama.dz; bouziane@cuniv-naama.dz; bouchiha@cuniv-naama.dz
Abstract:
This paper presents a deep learning approach for Arabic question classifica-tion, leveraging the capabilities of the AraELECTRA language model to generate word representations and a convolutional neural network for classi-fication. The dataset employed in this study, translated from English to Ara-bic, categorizes questions following the Li and Roth taxonomy. The results demonstrate an accuracy of 85.32%, showcasing the potential of this ap-proach to contribute significantly to the development of Arabic question-answering systems and support researchers in improving their accuracy.
Depth-Aware Topic Modeling for Expertise Detection in Social Bookmarking Platforms
Authors: Saida Kichou (CERIST)*; Fouad Dahak (ENSIA); Abdelkrim Meziane (CERIST)
Emails: saidakichou@gmail.com; fouad.dahak@ensia.edu.dz; ameziane@cerist.dz
Abstract:
Identifying user expertise through digital traces has become a key challenge in social computing and user profiling. Social bookmarking platforms, where users freely annotate and organize content with tags, offer a valuable source for assessing individual knowledge domains. In this paper, we extend a previous approach to expertise evaluation by integrating tag depth into the Latent Dirichlet Allocation (LDA) topic modeling process. Our method leverages both the content and the hierarchical structure of tags to enhance topic representation and better capture users' actual areas of expertise. The approach is applied to data from the Delicious platform, where we analyze tagging behaviors to infer expertise profiles. Experimental results show that incorporating tag depth improves topic specificity and provides more meaningful, quantifiable indicators of expertise. This work highlights the potential of semantic tag structures in refining topic modeling and supports the use of social tagging systems as a reliable basis for expert identification.
NEXT WORD PREDICTOR USING LSTM
Authors: vivek kharate (vishwakarma institute of information tecnology)*; jyotiraditya chvan (vishwakarma institute of information tecnology); ratna patil (vishwakarma institute of information tecnology)
Emails: vivek.22210158@viit.ac.in; jyotiraditya.22210606@viit.ac.in; ratna.patil@viit.ac.in
Abstract:
Long words tire to type, but predictive text software in keyboards simplifies it. Another name for next-word prediction is language modelling. One's work is simply predicting the first word to be spoken. It has a number of applications and constitutes the basic human language technology work. This technique is letter to letter prediction and it claims that it is predicting a letter when word is constructed in terms of letter. Long short time memory formula can sense previously typed text and predict words which can be repeated again for people to enclose sentences.
Speaker Verification Using Multi-Scale K-Neighboring Residual network for Robust Embedding Learning
Authors: Abderrachid Bouzebra (Brothers Mentouri University)*; Atef Farrouki (Brothers Mentouri University)
Emails: bouzebraabderrachid@gmail.com; farroukiatef@yahoo.fr
Abstract:
We propose a novel convolutional neural network (CNN) architecture for speaker verification, designed to effectively capture sequential and local interactions within input speech mel-filterbanks. Our approach integrates a series of multi-scale K-neighboring residual convolutional (MKRC) blocks, which enable each sub-feature to integrate contextual information from its local neighbors. This architecture facilitates the generation of speaker embeddings with an enhanced ability to differentiate between similar speakers. At first, Mel-filterbanks are extracted from each input speech, which is fed to the proposed model for speaker embedding generation. Euclidean distance is calculated between pairwise embeddings for performance evaluation. The proposed model has been compared with state-of-the-art methods using VoxCeleb1 dataset. Experimental results show that our model achives promising results in terms of Equal Error Rate (EER) and minimum Detection Cost Function (minDCF).
Anomaly Detection and Diagnosis of Wind Turbines Using Deep Learning Techniques: Aeolian Wind Speed Case Study
Authors: BRAHAMI Menaouer (National Polytechnic School of Oran - Maurice Audin)*; Bezzemmit Cha ma (National Polytechnic School of Oran - Maurice Audin); Sabri Mohammed (National Polytechnic School of Oran - Maurice Audin); Matta Nada (university of technology of troyes)
Emails: brahami.menaouer@gmail.com; chaimaabezzemmit@gmail.com; ram.sabri@gmail.com; nada.matta@utt.fr
Abstract:
In recent years, wind turbine condition monitoring based on Supervisory Control and Data Acquisition (SCADA) systems has attracted considerable scientific research interest. Frequently reported challenges include the fact that most wind turbine SCADA parameters are highly dependent on the operating conditions, such as wind speed, wind direction, and LV Active Power, along with the control actions imposed on the wind turbine. Thus, strict and effective data quality control of the SCADA data is crucial. Besides that, intelligent anomaly detection for wind turbines using artificial intelligence techniques has been extensively researched and yielded significant results. Likewise, the usage of machine/deep learning techniques is widely spread and has been implemented in the wind industry in the last few years. The development of sophisticated deep learning now allows improvements in anomaly detection from historical data. In this paper, we present a wind speed anomaly detection approach using LSTM (Long Short-Term Memory), CNN (Convolutional Neural Network), and GRU (Gated Recurrent Unit) to detect the minimum and maximum values of wind speed. The approach applies the information in supervisory control and data acquisition systems of Aeolian wind speed. This comprehensive approach offers a promising avenue for the precise anomaly detection of Aeolian wind speed, providing practicians with a reliable tool for accurate diagnosis, critical for timely intervention. Real cases from a wind farm have confirmed the feasibility and advancement of the proposed Deep Learning models, while also discussing the effects of various applied parameters.
Speech Emotion Recognition Based on Gender Influence in Different Languages Using Various Classifiers
Authors: Horkous Houari (Center for Development of Advanced Technologies)*
Emails: houarihorkous29@yahoo.fr
Abstract:
Speech emotion recognition (SER) is an attractive and challenging task in human-computer interaction and artificial intelligence technologies. SER is the process of recognizing emotions from speech utterances. This work is based on emotion recogni-tion from speech, and our purpose is to study the effect of gender classes on the SER. Three emotional databases with different languages: ADED, EMO-DB and ShEMO in Algerian dialect, German and Persian languages respectively, are used to evaluate the performance of the SER system. Extracting features from speech is a required step in the creation of the SER system. Combinations of prosodic and MFCC features are exploited as speech features in the system of recognition. This last is based on Linear Discriminate Analysis (LDA), Deep Neural Networks (DNN) and Support Vectors Machine (SVM) as methods of classification. The results obtained show us that the system of SER is influ-enced by gender classes. The recognition rates of SER systems with gender distinction are higher than the recognition rates of SER systems without gender distinction.
FsMo- FsAe: A Comparative Study of Multi-Objective Using Binary Differential Evolution and Autoencoder-Based Feature Selection on the Lung Gene Expression Data
Authors: Mohamed DJELLAL SERANDI (UNIVERSITY OF MASCARA)*; Fatma BOUFERA (UNIVERSITY OF MASCARA); Amina HOUARI (UNIVERSITY OF MASCARA ); Farid FLITTI (Higher Colleges of Technology, Campus Dubai, UAE)
Emails: mohamed.djellalserandi@univ-mascara.dz; fboufera@univ-mascara.dz; amina.houari@univ-mascara.dz; flitti@hct.ac.ae
Abstract:
High-dimensional datasets, such as those related to lung gene expression, present major challenges due to the presence of irrelevant or redundant features. In this study, we conduct a comparative analysis of multi-objective feature selection and an autoencoder-based approach to address this issue. We employ the Binary Differential Evolution (BDE) algorithm and evaluate three configurations: the original dataset (without feature selection), an autoencoder-based feature selection method(FsAe), and a multi-objective feature selection method (FsMo), which simultaneously optimizes the Mean Squared Residue (MSR) and the number of selected features. Our experimental results show that the FsMo method outperforms both the autoencoder-based method and the unfiltered dataset in terms of classification accuracy.
Development and Implementation of an AI-Driven System for Pancreatic Cancer Susceptibility Prediction, Diagnosis, and Survivability Estimation Using Machine Learning Algorithms.
Authors: Amina BOUMEDIENE (Universite Oran1 Ahmed ben Bella, Oran, Algerie)*; Asmaa BENGUEDDACH (Universite Oran1 Ahmed ben Bella, Oran, Algerie); Karim BOUAMRANE (Universite Oran1 Ahmed ben Bella, Oran, Algerie)
Emails: boumediene_mn@outlook.fr; asmaa.bengueddach@gmail.com; kbouamrane20@gmail.com
Abstract:
Pancreatic cancer is one of the most aggressive and lethal malignancies, with a five-year survival rate remaining alarmingly low due to late-stage diagnosis and limited treatment options. Early detection and accurate prognostic assessment are essential for improving patient outcomes, yet traditional diagnostic methods rely heavily on invasive procedures and subjective clinical assessments. Recent advances in artificial intelligence (AI) and machine learning (ML) have demonstrated significant potential in enhancing cancer detection and prediction through data-driven, automated approaches. However, existing AI models often focus on a single predictive task or require predefined task selection, limiting their adaptability in real-world clinical applications. This study aims to address these limitations by developing an intelligent system capable of automatically identifying and executing the appropriate predictive tasknamely, susceptibility prediction, diagnosis, or survivability estimationbased on the input dataset.
A Comparative Analysis of Data Mining Approaches for Phishing Email and Website Detection
Authors: Janish Macwan (Marwadi University )*; Dr. Jaypalsinh Gohil (Marwadi University)
Emails: janish.macwan120547@marwadiuniversity.ac.in; jaypalsinh.gohil@marwadieducation.edu.indieducation
Abstract:
Targeting people and companies to compromise private data, phishing is a ubiquitous cybercrime. Conventional methods of detection have shown inadequate ability to stop developing phishing campaigns. Emphasizing the importance of data mining (DM) techniques, this work offers a methodical and comparative review of phishing website and email detection systems. Analyzed are several DM techniques, datasets, feature engineering approaches, and evaluation metrics used in recent work. This review notes important trends, points up current research gaps, and suggests future directions to improve detection efficiency and robustness. Our results seek to be a complete source of reference for cybersecurity practitioners and academics. This work reveals that compared to standalone models, ensemble data mining models exhibit better adaptation to changing phishing strategies. (Abdelhamid, 2022)
Optimizing Human Resource Allocation in Emergency Departments: A Combined R-NSGA-III and Colored Petri Nets Approach
Authors: zouaoui louhab (Department of Computer Science, University Mustapha Stambouli, Mascara, Algeria)*; fatma boufera (LISYS Lab, University Mustapha Stambouli, Mascara, Algeria ); boudjelal meftah (LISYS Lab, University Mustapha Stambouli, Mascara, Algeria)
Emails: zouaoui.louhab@univ-mascara.dz; fboufera@univ-mascara.dz; boudjelal.meftah@univ-mascara.dz
Abstract:
Recently, severe overcrowding in the emergency department has exacerbated the challenges faced by both patients and staff. This overcrowding leads to the pa-tient's long stay in the emergency department as well as the financial losses of the hospital. Emergency department are vital in society, where the patient comes to it at any time, and without a prior date. The emergency department is a complex system due to the nature of the available resources in addition to unpredictable cases. Recently, many researchers have focused on reducing the time patients spend in the ED to alleviate pressure on medical staff and improve the quality of services offered to patients. In this paper, we propose an approach that combines R-NSGA-III algorithms with colored Petri nets. The emergency department is modeled using colored Petri nets, and initial results are obtained after running the simulation model. R-NSGA-III algorithms are relied on to obtain the optimal number of human resources, so that the simulation model is modified with its re-starting each time and compares the results with the various proposed models. This approach helps the hospital decision-makers to find effective solutions, es-pecially in terms of human resources.
A Review on Cryptocurrency Fraud Detection Techniques: Challenges, Solutions, and Perspectives
Authors: BOUCHAMA Salah-eddine (Universit Blida 1, Laboratoire LRDSI, Facult des Sciences, BP 270, Route de Soumaa, Blida, Alg rie)*; Samir OUCHANI (CESI LINEACT, Aix-en-Provence); Hafida BOUARFA (Universit\'e Blida 1, Laboratoire LRDSI, Facult\'e des Sciences, BP 270, Route de Soumaa, Blida, Alg rie)
Emails: bouchama1976@gmail.com; souchani@cesi.fr; hafi.bouarfa@gmail.com
Abstract:
The rapid expansion of fraudulent behavior concerning the cryptocurrency ecosystem has underscored the necessity for intelligent detection frameworks. This paper focuses on a comparative review of modern approaches organized into three methodological categories: cuttingedge AI frameworks, traditional ML methods, and non-ML or heuristic approaches. In analyzing sixteen contributions, we assess the multidisciplinary gaps and challenges of each class in terms of model interpretability, data imbalance, high computational cost, overfitting, insufficient empirical evaluation, and generalizability to novel instances of fraud. To overcome these challenges, we propose the use of explainable AI tools such as SHAP and GNNExplainer, along with hybrid detection pipelinescombining deep learning, traditional models, and heuristic knowledgedata generation through transfer learning, and selfsupervised frameworks. We also highlight the need for modular designs focused on system scalability and evolvability, as well as the integration of on-chain and off-chain data to improve situational awareness. Given the complexity of cryptocurrency fraud, we advocate for strategically adaptable, explainable, and ethically responsible detection frameworks that enhance performance and explainability in the ever-changing environment of decentralized blockchain systems. Keywords: Cryptocurrency Fraud Artificial Intelligence Machine Learning Deep Learning Graph Neural Networks Hybrid Methodologies Explainable AI Heuristic Approaches.
ROI Detection and Multi-Classification of Skin Lesion Using Deep Learning techniques
Authors: Amel Ykhlef (Ecole nationale polytechnique d'Oran)*
Emails: amelykhlef13@gmail.com
Abstract:
Skin cancer is one of the most dangerous diseases in the world and accounts for the highest number of deaths each year. According to the Skin Cancer Foundation, the number of newly diagnosed melanoma cases in the United States is expected to increase by 7.3% in 2024, and the number of melanoma deaths is expected to increase by 3.8%, This is why artificial intelligence techniques have been used to solve skin cancer problems, using deep learning techniques for classification, detection and prediction. In this work, we present a deep learning technique for object detection, known as YOLO (You Only Look Once ) as a first step rather than using convolutional neural network (CNN) for multi-classification. The results showed that the model achieved accuracy of 94% for testing data and 98 % for training data. This methodology shows that the use of object detection techniques can improve classification performance.
NeuroVisionNet: A Modified EfficientNetB3 Frame-work with Grad-CAM for Multi-Class Brain Tumor MRI Classification
Authors: Souhila AIT ABDERRAHIM (National Polytechnic School of Oran)*; Menaouer BRAHAMI (National Polytechnic School of Oran); Abdelmalek AMINE (University of Saida - Dr. Moulay Tahar)
Emails: souhila.aitabderrahim@gmail.com; Menaouer.brahami@enp-oran.dz; amine_abd1@yahoo.fr
Abstract:
Brain tumors represent one of the most critical neurological diseases due to their heterogeneous morphology and potential for severe clinical outcomes. Accurate and timely classification is essential to guide therapeutic decisions and improve patient prognosis. In this study, we propose NeuroVisionNet, a deep learning framework designed for multi-class classification of brain tumors glioma, meningioma, pituitary tumor, and no-tumor cases using contrast-enhanced T1-weighted MRI scans. The model is based on a modified EfficientNetB3 architecture, optimized through transfer learning, fine-tuning, and regularization strategies. To enhance interpretability, we integrate Grad-CAM visualizations, enabling clinicians to inspect salient tumor regions influencing predictions. Experimental evaluation on the publicly available Kaggle brain tumor MRI dataset (~3,000 images) demonstrates that NeuroVisionNet achieves 98.73% classification accuracy, with class-wise precision and recall consistently above 97%. We provide a detailed per-class performance analysis, including clinical implications, and compare our model against established baselines such as ResNet50 and DenseNet121. While our study is limited by dataset size and single-source origin, the results highlight the feasibility of deploying such lightweight, explainable models in clinical workflows, including potential integration with PACS systems (Picture Archiving and Communication System).
A Historical Overview of Image Quality Assessment Methods: Focus on Medical Imaging Applications
Authors: ardjani fatima (University Center Nour Bachir El Bayadh, Algeria); KHALDI Abdelkrim (University Center Nour Bachir El Bayadh, Algeria)*; TAOUAF Lakhdar (University Center Nour Bachir El Bayadh, Algeria); BOUCHIHA Djelloul (University Centre -Salhi Ahmed- Naama)
Emails: f.ardjani@cu-elbayadh.dz; a.khaldi@cu-elbayadh.dz; l.taouaf@cu-elbayadh.dz; bouchiha@cuniv-naama.dz
Abstract:
This paper provides a comprehensive review of Image Quality Assessment (IQA) methods, tracing their historical development from early conventional met-rics to modern deep learning-based approaches. First, it describes fundamental subjective and objective techniques, including full-reference, reduced-reference, and no-reference methods. Then, the review examines major advancements across different eras, including HVS-based models, transform-domain and natu-ral scene statistics techniques, and traditional machine learning approaches. Spe-cial attention is given to recent deep learning innovations, particularly convolu-tional neural networks (CNNs), vision transformers, and modern training para-digms such as transfer learning, meta-learning, and self-supervised learning. The survey emphasizes the applications in medical imaging, where accurate and ro-bust IQA is critical for reliable diagnosis and clinical decision-making. Finally, the paper highlights ongoing challenges and outlines future research directions for building medically effective and reliable IQA systems. This paper aims at serving as a comprehensive reference for researchers seeking to understand exist-ing approaches, identify limitations, and develop new solutions tailored to do-main-specific needs.
Advancing Traffic Management: A Real-Time CNN-Based Multi-Vehicles Speed Estimation and License Plate Recognition
Authors: Selsabil Benderradji (University of kasdi merbah ouargla)*; Chahira Kezzal (University of kasdi merbah ouargla); Azeddine Benlamoudi (University of kasdi merbah ouargla); Salah Eddine Bekhouche (San Sebastian, Spain); Marouane Dida (University of kasdi merbah ouargla); Abdelhai Lati (University of kasdi merbah ouargla)
Emails: benderradji.selsabil@univ-ouargla.dz; kezzal.chahira@univ-ouargla.dz; benlamoudi.azeddine@univ-ouargla.dz; sbekhouche001@ikasle.ehu.eus; dida.marouane@univ-ouargla.dz; lati.abdelhai@univ-ouargla.dz
Abstract:
The growing complexity of modern traffic systems calls for robust, real-time solutions for vehicle monitoring and enforcement. This paper presents an integrated framework for multi-vehicle speed estimation and license plate recognition (LPR), combining the strengths of three state-of-the-art technologies: YOLOv11 for object detection, SORT for multi-object tracking, and EasyOCR for optical character recognition. The system detects vehicles and license plates in real-time video streams,tracks their movement, estimates vehicle speed using pixel displacement and temporal data, and recognizes license plate characters with high accuracy. Experimental results show that YOLOv11 achieves superior performance compared to YOLOv8 and YOLOv9, with a precision of 94.5%, recall of 75.9%, F1-score of 84.2%, and mAP@0.5 of 85.6%. The model demonstrates robust detection even under challenging conditions such as occlusion, Perspective Distortion, and plate angle variation. The results validate the model's potential for integration into Algerian national traffic management systems and future smart city infrastructure.
Examining Privacy-Utility Tradeoffs in Differentially Private Medical Image Classification with Data Augmentation
Authors: Benladghem Rafika (LRIT- Tlemcen Univesity)*; Hadjila Fethallah (LRIT- Tlemcen Univesity); Belloum Adam (Informatics Institute, University of Amsterdam, Amsterdam, The Netherlands)
Emails: rafika.ben.2492@gmail.com; hadjila.fethallah@gmail.com; a.s.z.belloum@uva.nl
Abstract:
Privacy protection in medical AI presents fundamental chal- lenges as healthcare datasets contain highly sensitive patient information subject to strict regulatory requirements. While differential privacy offers rigorous mathematical guarantees for privacy-preserving machine learning, it typically reduces model performance through noise injection. Simultaneously, data augmentation addresses critical challenges in medical imaging including limited training data and class imbalances. However, the interaction between these widely-used techniques remains unexplored, creating uncertainty for practitioners implementing privacy-preserving medical AI systems. This paper presents a systematic empirical study examining how data augmentation affects privacy-utility tradeoffs in medical image classification. Using the PneumoniaMNIST dataset for pneumonia detection, we evaluate rotation-based augmentation combined with differentially private training across privacy budgets ranging from = 1.0 to = 8.0. Our comprehensive experiments reveal complex non-linear relationships between privacy parameters, augmentation strategies, and model performance. Key findings demonstrate that moderate privacy budgets ( = 8.0) with rotation augmentation achieve optimal balance, maintaining 83.8% accuracy while providing meaning-ful privacy protection. We identify a critical "privacy cliff" below = 1.0 where utility becomes clinically unacceptable (62.5% accuracy), establishing practical lower bounds for medical AI applications. Results show that augmentation-privacy interactions are context-dependent, with augmentation improving baseline performance by 1.2% but yielding mixed results when combined with privacy mechanisms. These findings provide evidence-based guidance for healthcare practitioners and policymakers balancing privacy protection with diagnostic accuracy, establishing practical privacy budget ranges for medical AI applications.
Duck Swarm Algorithm-Based Clustering Technique
Authors: Ibrahim Zebiri (Universit 20 Ao t 1955)*; Nor Samsiah Sani (Universiti Kebangsaan Malaysia)
Emails: ibrahimzebiri97@gmail.com; norsamsiahsani@ukm.edu.my
Abstract:
Data clustering is an unsupervised task that aims to subdivide a set of unlabeled data into a number of homogeneous groups, it is used in several scientific fields such as bioinformatics, social sciences, psychology, chemistry, materials science, medicine and healthcare. A central challenge to data clustering is verifying all possible solutions to find the best one, which is beyond our capacities for small values of instances and clusters, not to mention that most of clustering applications come with quite bigger parameters. Thus, an effective technique is needed that can be employed with usual and large datasets. This work presents an adaptation of a recent metaheuristic called Duck Swarm Algorithm (DSA) in order to tackle data clustering problem. The adapted version (DSAC) is compared to different well-known and recent algorithms and tested on several real clustering datasets to reveal its performance. Experimental results exposed the superiority of the proposed DSAC in finding optimal clusters.
Agentic Retrieval-Augmented Generation for Arabic Legal Data
Authors: zoubida boudjenane (University of Mascara)*; Mohammed Salem (University of Mascara)
Emails: zoubida.boudjenane@univ-mascara.dz; salem@univ-mascara.dz
Abstract:
Legal texts are often long, unstructured, and domain-specific, posing challenges for traditional NLP systemsespecially in low-resource languages like Arabic. This paper introduces an Agentic Retrieval Augmented Generation (RAG) system tailored for Arabic legal question answering, with a focus on rulings from the Algerian Supreme Court. By combining document-grounded generation with modular, goal-directed behavior, the system enhances retrieval quality and response reliability. Unlike conventional RAG, our agent includes specialized components for filtering, recommending legal articles, and generating answers grounded in real legal sources. To support this system, we develop two Arabic legal datasets: annotated rulings and synthetic questionanswer pairs. Experimental results show that our agent consistently retrieves relevant context and generates accurate, fluent Arabic answers. This work demonstrates how integrating agentic reasoning with RAG can improve transparency and factual accuracy in legal NLP, offering a robust solution for complex legal queries in low-resource environments.
Using Machine Learning to Predict Behavioral Effects of Social Media Reels on Young Users
Authors: laouni mahmoudi (university mustapha stambouli of mascara)*; Mustapha Sahraoui (university mustapha stambouli of mascara); Abbes Belgoumidi (Department of Psychology, University Oran 2, Oran 31000, Algeria); mohammed abdelatif djebbar (Department of Psychology, University Oran 2, Oran 31000, Algeria)
Emails: laouni.mahmoudi@univ-mascara.dz; sahraoui.musta@univ-mascara.dz; absbelgomidi@yahoo.fr; djebbar.med.abdelatif@gmail.com
Abstract:
This study investigates the influence of AI-Powered video recommendations, such as Facebook Reels and TikTok videos, on young users' information-seeking behaviors and potential addiction. By integrating machine learning and psychological insights, it aims to highlight how these recommendations can unintentionally lead to prolonged engagement and increased screen time. Using a dataset of 1,000 young Algerians, the study analyzes interactions with AI-generated Reels, employing machine learning models specifically RandomForest and GradientBoosting to identify patterns in engagement influenced by AI suggestions. Collaboration with psychological researchers aids in detecting correlations between features such as Reel Session Duration, Number of AI-Suggested Reels, age, and education level, and their association with addictive behaviors. Findings indicate that AI recommendations significantly enhance user engagement, with the RandomForest model achieving 74\% accuracy and an F1 Score of 0.783. The GradientBoosting model performed slightly better, with 75\% accuracy and an F1 Score of 0.805. This study underscores the profound impact of AI recommendations on young users and highlights the need for user-centric algorithm design in social media, as responsible AI systems can enhance user experiences while prioritizing mental well-being.
An Advanced Security Model for Virtualized Network Functions in the Cloud
Authors: REBBAH MAROUA (Department of Computer Science, University of Mustapha Stambouli); REBBAH MOHAMMED (Department of Computer Science, University of Mustapha Stambouli)*; SMAIL OMAR (Department of Computer Science, University of Mustapha Stambouli); DEBAKLA MOHAMMED (Department of Computer Science, University of Mustapha Stambouli)
Emails: rebbahmarwa13@gmail.com; rebbahmed@univ-mascara.dz; o.smail@univ-mascara.dz; debakla_med@univ-mascara.dz
Abstract:
Cloud computing and virtualization have revolutionized the IT industry, providing scalability and flexibility. But the dynamic and dispersed character of NFV-based cloud systems presents serious security issues including vul-nerabilities in virtualized network functions (VNFs), inter-VNF communication, and hypervisor attacks. This paper focuses on the security implications of network virtualization and proposes a novel security model based on artificial intelligence and deep learning algorithms. The model combines a Deep Autoen-coder (DAE) and machine learning techniques for network intrusion detection in NFV environments. The results demonstrate the model's effectiveness in de-tecting network intrusions in virtual networks. Our model achieved a high de-tection rate of 96%, demonstrating its effectiveness in detecting network intru-sions within virtual networks.
Optimal Resource Provisioning for Scientific Workflow Ensembles in Cloud Computing
Authors: REBBAH MOHAMMED (Department of Computer Science, University of Mustapha Stambouli)*; BENATIA Salah Eddine (Department of Computer Science, University of Mustapha Stambouli); SMAIL OMAR (Department of Computer Science, University of Mustapha Stambouli); DEBAKLA MOHAMMED (Department of Computer Science, University of Mustapha Stambouli)
Emails: rebbahmed@univ-mascara.dz; se.benatia@univ-mascara.dz; o.smail@univ-mascara.dz; debakla_med@univ-mascara.dz
Abstract:
A workflow ensemble consists of multiple sub-workflows, each containing hundreds or thousands of jobs with constrained priorities. Their implementations are widely accepted. Furthermore, extensive research and development work has been done to automate them and make them effective in cloud computing environments. This paper addresses the challenge of resource provisioning for large-scale scientific workflows running on many heterogeneous clouds by running all processes of the ensemble under budget and time constraints. We have de-veloped four algorithms for resource provisioning and job scheduling based on static and dynamic methods. Si-mulations based on real scientific workflow applications are applied in experimental evaluation. The results show the ability of the proposed approach of reducing budgets and execution times for various workflow ensembles.
Efficient Breast Cancer Identification Using Sunflower Optimization-Based Feature Selection with Multiple Classifiers
Authors: Mohamed Debakla (Universit de Mascara)*; Ali Mezaghrani (University of Mustapha Stambouli, Mascara); Khalifa Djemal (University of Evry Paris-Saclay); Mohammed Rebbah (University of Mustapha Stambouli, Mascara)
Emails: debakla_med@univ-mascara.dz; Ali.mezaghrani@univ-mascara.dz; khalifa.djemal@ibisc.univ-evry.fr; rebbahmed@univ-mascara.dz
Abstract:
Breast cancer is one of the leading causes of mortality among women worldwide, and its early and accurate diagnosis is critical for effective treatment. In this study, we propose a hybrid classification framework that combines Feature Selection (FS) using the Sunflower Optimization (SFO) algorithm with multiple classification models including Support Vector Machine (SVM), Random Forest (RF), Na ve Bayes (NB), and Convolutional Neural Networks (CNN). The SFO algorithm, inspired by the phototropic behavior of sunflowers, is employed to identify the most informative subset of features from the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, thereby reducing dimensionality and enhancing classifier performance. Each selected feature subset is then evaluated using classical machine learning classifiers (SVM, RF, NB) and a deep learning model (CNN) to compare classification accuracy and generalization capability. The models are assessed based on performance metrics including accuracy, precision, sensitivity, specificity, and F1-score. Experimental results demonstrate that the SFO-based feature selection significantly improves the predictive accuracy across all classifiers, with CNN achieving the highest performance, reaching an accuracy exceeding 98,64%. The proposed approach proves to be an effective, robust, and interpretable solution for breast cancer classification tasks, especially in scenarios where dimensionality reduction and classification reliability are essential.
Knowledge Distillation of Vision Transformers for Multiple sclerosis Lesion Classification in Brain MRI imaging
Authors: LARIBI NOR-ELHOUDA (University M'Hamed Bougara Boumerd s, Algeria)*; GACEB Djamel (University M'Hamed Bougara Boumerd s, Algeria ); REZOUG Abdellah (University M'Hamed Bougara Boumerd s, Algeria ); TOUAZI Fay al (University M'Hamed Bougara Boumerd s, Algeria )
Emails: n.laribi@univ-boumerdes.dz; d.gaceb@univ-boumerdes.dz; a.rezoug@univ-boumerdes.dz; f.touazi@univ-boumerdes.dz
Abstract:
Multiple sclerosis lesion detection in brain MRI remains a challenging task due to lesion hetero-geneity, class imbalance, and variability in imaging protocols to detect the progression of this lesion. In this work, we present a new study to apply knowledge distillation using knowledge distillation with vision transformers for Multiple sclerosis lesion detection in brain MRI, focusing on transferring knowledge from a powerful Pyramid Vision Transformer teacher model to a lightweight MobileViT student model. Experimental results show that Pyramid Vision Trans-former, as the most performant teacher, significantly, increasing the accuracy of student models from 91.25% to 94.64% and F1-score from 89.4% to 93.19%, achieving a 4% gain on a limited dataset. This study shows that using a powerful model like PVT as a teacher in a knowledge distillation framework can effectively improve the performance of smaller models such as Mo-bileViT, even when training data is limited or imbalanced. By transferring rich feature representa-tions, the approach enables lightweight models to achieve high accuracy and generalization, mak-ing them suitable for deployment in resource-constrained healthcare settings.
Comprehensive Assessment of Neuromuscular Dysfunction in Parkinson's Disease Using Multi-Axial Accelerometers
Authors: Touahria Rima (LMSE laboratory University Mohamed El-Bachir El-Ibrahimi bordj Bou Arreridj Algeria )*; Hacine Gharbi Abdenour (LMSE laboratory University Mohamed El-Bachir El-Ibrahimi bordj Bou Arreridj Algeria); Ravier Philippe (PRISME laboratory, University of Orleans, 12 rue de Blois, 45067 Orleans, France); Abed-Meraime Karim (PRISME laboratory, University of Orleans, 12 rue de Blois, 45067 Orleans); Buttelli Olivier (PRISME laboratory, University of Orleans, 12 rue de Blois, 45067 Orleans)
Emails: touahria.rimaa@gmail.com; gharbi07@yahoo.fr; philippe.ravie@univ-orleans.fr; karim.abed-meraim@univ-orleans.fr; olivier.buttelli@univ-orleans.fr
Abstract:
This study investigates the potential of multi-axial accelerometers (ACC) for a comprehensive assessment of neuromuscular dysfunction in Parkinson's disease (PD) patients. Emphasizing the importance of utilizing data from three axes (ACC_X,ACC_Y,and ACC_Z) to discriminate the movement characteristics, the proposed method leverages Wavelet Cepstral Coefficients (WCC) extracted from ACC data. A k-Nearest Neighbour (KNN) classifier is employed to differentiate between normal and Parkinson. Feature selection based on mutual information criteria further optimizes the classification process in terms of complexity and accuracy. Evaluation using the ECOTECH project database yielded a promising accuracy of vectors of 98.91% with 100% of signals classification rate(CRs) and F_score of 98.96% using only two features selected using the Conditional Infomax Feature Extraction filter selection strategy were sufficient to explain the two classes (N) and (P). These findings suggest that ACC_Y data combined with WCC feature extraction and KNN classification holds promise as efficient tool for PD diagnosis and potentially monitoring disease progression.
Improving Dialectal Text Classification via Attention-Based Stopword Extraction for Algerian Arabic
Authors: Lamia Ouchene (Ferhat Abbas Setif 1 University)*; Sadik Bessou (University of Ferhat Abbas Setif 1)
Emails: lamia.ouchene@univ-setif.dz; bessou.s@univ-setif.dz
Abstract:
In low-resource dialects like the Algerian dialect, traditional stopword lists are often incomplete and poorly suited to informal text such as social media content. To address this challenge, we present a novel method for constructing a stopword list tailored to dialectal Arabic. The method leverages model interpretability techniques. Building on the self-attention mechanism of MARBERT, a transformer model trained on Arabic and its dialects, we identify low-utility words based on their attention patterns during sentiment classification. We track words that attract disproportionately high attention in misclassified samples but low attention in correct predictions. Such words may serve as sources of noise rather than informative features. Candidate stopwords are dynamically evaluated through a probabilistic removal process during training. Experiments on a manually annotated Algerian dialect dataset demonstrate the effectiveness of the method. The attention-based stopword extraction improves classification accuracy over static stopword models. This approach offers a practical solution for building effective stopword lists in low-resource dialects.
Neuro-Fuzzy Approach for Nonlinear System Control
Authors: DAIKH fatima zohra (University Mustapha Stambouli of Mascara)*
Emails: fatima_daikh@yahoo.fr
Abstract:
In this article, we address the exploitation of Artificial Intelligence properties in the field of control engineering. Our work focuses on the use of neuro-fuzzy networks, specifically ANFIS (Adaptive Neuro-Fuzzy Inference Sys-tem) and STFIS, for the identification of models required to design control laws for a nonlinear dynamic system the inverted pendulum. In the first part, the ANFIS model is used as a controller in multiple struc-tures. The second section presents an application of the STFIS controller on the nonlinear system. The objective of this paper is to improve the perfor-mance of the ANFIS and STFIS models for a system subjected to a constant disturbance. The proposed approach is validated through simulations carried out in the MATLAB environment.
Hybridization CNN and Fuzzy Logic approch for the Identification of Intraocular Eye Disease
Authors: Sahraoui Mustapha (University of MASCARA)*; Mahmoudi Laouni (University of MASCARA); Nawaf Alharbe (Department of Computer Scienceand Engineering, Madinah, Saudi Arabia)
Emails: sahraoui.musta@gmail.com; laouni.mahmoudi@univ-mascara.dz; nrharbe@taibahu.edu.sa
Abstract:
Ocular hypertonia, a major risk factor for glaucoma, leads to gradual retinal degradation and is often asymptomatic in its early stages, making early detection challenging. In recent years, the use of Artificial Intelligence (AI), particularly deep learning, has significantly enhanced the analysis and interpretation of medical images for diagnostic purposes. This project aims to detect and classify ocular hypertension through the analysis of fundus images using deep learning. A hybrid approach combining a Convolutional Neural Network (U-Net) with fuzzy logic known as a Neuro-Fuzzy system was employed. The U-Net model automatically extracts features from preprocessed fundus images and classifies them into two categories: diseased or normal. Fuzzy logic is then applied to refine this classification into three stages, improving diagnostic precision. The integration of U-Net with fuzzy logic has demonstrated strong potential in retinal pathology assessment, offering improved handling of uncertainty and variability inherent in medical data.
Artificial Intelligence Techniques for Cyberbullying Detection: A Comprehensive Review of Machine Learning and Deep Learning models
Authors: kaddaouia Habib (Universit de mascara)*; Fatima Debbat (Universit de mascara); Mohammed Salem (universit se mascara)
Emails: kadaouia.habib@univ-mascara.dz; debbatfatima@gmail.com; salemohammed@gmail.com
Abstract:
Cyberbullying has emerged as a critical threat in online social platforms, particularly among adolescents and vulnerable users. The rapid growth of internet and social media user's generated content on social media has intensified the need for automated systems capable of detecting harmful content. This survey provides a comprehensive overview of recent advancements in cyberbullying detection using traditional machine learning (ML)methods such as SVM, NB and AdaBoost, deep learning (DL) models like CNN and transformers, and hybrid approaches that combine ML and DL models. We review a wide range of studies published between 2023 and 2025, highlighting the evolution of used techniques. Traditional ML methods have given high performance for text classification. However, DL models like CNNs, LSTMs and transformers have significantly improved performance by capturing contextual and semantic patterns in textual data. Moreover, hybrid systems that integrate ML and DL are increasingly being adopted to combine the strengths of both techniques. The survey also discusses datasets used for training and evaluating models. This work aims to guide future research for more pertinent and robust solutions for cyberbullying detection.
A New Fuzzy Clustering Validity Index Based on Kullback-Leibler Divergence
Authors: chakir mokhtari (mascara university)*; mohammed debakla (mascara university ); boudjelal meftah (mascara university )
Emails: chakir_m@yahoo.fr; debakla_med@univ-mascara.dz; boudjelal.meftah@univ-mascara.dz
Abstract:
Determining the optimal number of clusters, a critical step in clustering analysis, is typically guided by domain expertise or assessed through clustering validity indexes. This study evaluates the effectiveness of such indexes for centroid-based partitional clustering algorithm. We propose a new clustering validity index, termed KLDCVI, which mitigates instability by incorporating Kullback-Leibler Divergence. This adjustment allows KLDCVI to tolerate closely allocated centroids to a reasonable degree, enhancing robustness in evaluation. We assess the performance of KLDCVI against existing validity indexes by applying the fuzzy c-means algorithm to real-world images. Experimental results demonstrate that KLDCVI achieves superior accuracy and reliability compared to conventional indexes.
Enhancing Grey Wolf Optimizer for Imbalanced Feature Selection in Diabetes Prediction via Chaotic Initialization
Authors: MOHAMMED DJAOUTI SOUMIA (UNIVERSITY OF MASCARA)*; BIDI NOURIA (UNIVERSITY OF MASCARA )
Emails: djouti2012@YAHOO.FR; n_bidi@univ-mascara.dz
Abstract:
Feature selection is a crucial pre-processing step in machine learning, particularly for high-dimensional datasets and imbalanced classification problems. It aims to identify a minimal subset of relevant features that significantly improve model performance and interpretability. This article proposes an enhanced Grey Wolf Optimizer (GWO) for feature selection, integrating chaotic maps for population initialization. The proposed approach, termed GWO-CFI (GWO with Chaotic feature Initialization), is applied to a pre-processed and balanced diabetes prediction dataset, utilizing a Random Forest classifier as the evaluation model. Experimental results demonstrate that chaotic initialization, specifically using Logistic, Tent, and Sine maps, can lead to more diverse initial populations, po-tentially improving the exploration capabilities of GWO and yielding superior feature subsets compared to standard random initialization, as evidenced by enhanced classification metrics (F1-score) on balanced dataset.
A Review on Deep Q-Network-Based Traffic Signal Control
Authors: Romaissa Bendaikha (University Mostafe Ben Boulaid -Batna2-)*; Djalal hedjazi (University Mostafe Ben Boulaid -Batna2-)
Emails: romaissa.bendaikha@univ-batna2.dz; d.hedjazi@univ-batna2.dz
Abstract:
Managing traffic signals is a significant challenge for modern transportation systems, contributing to congestion and environmental degradation. Artificial intelligence (AI) techniques, including Reinforcement Learning (RL), and, more specifically, Deep Q-Networks (DQN), have demonstrated considerable potential in addressing these issues by enabling efficient adaptive traffic signal control (ATSC). This paper comprehensively reviews DQN-based approaches applied to ATSC, focusing on key performance metrics such as vehicle waiting time, queue lengths, and traffic throughput. We explore how DQN addresses these challenges, offering insights into its effectiveness in optimizing traffic signal management at intersections.
A Genetic-Algorithm-Based Optimized Routing Protocol in Mobile Ad Hoc Networks
Authors: mekkaoui abdelkader (univerty mustapha stambouli mascara)*
Emails: abdelkader.mekkaoui@univ-mascara.dz
Abstract:
Mobile ad hoc networks (MANETs) are decentralized wireless systems in which nodes communicate without any fixed infrastructure. Their dynamic topology, limited bandwidth, and energy constraints present significant challenges for achieving efficient and reliable routing. To address these issues, bio-inspired optimization techniques have attracted growing interest due to their adaptability and robustness in complex environments. In this paper, we propose a multipath routing protocol based on a bio-inspired Genetic Algorithm (GARP). The multiple paths generated by the route discovery mechanism are optimized using the genetic algorithm to identify the most efficient path to the destination node. The path with the highest routing value evaluated based on node energy and link stability constraints is selected as the optimal solution. The proposed protocol is compared to existing approaches, namely AOMDV and AODV. To assess its performance, several key metrics are considered, including packet delivery ratio, routing overhead and energy consumption.
Arabic Intent Classification: From Data Creation to Prediction
Authors: Djahid DERBALE (University Centre of Na ma); Abdelghani BOUZIANE (University Centre of Na ma); Djelloul BOUCHIHA (University Centre of Na ma)*; Mohammed Amine BOUCETTA (University Centre of Na ma); Abdelkrim BERRABAH (University Centre of Na ma); Benamar HAMZAOUI (University Centre of Na ma)
Emails: jahiddr02@gmail.com; bouziane@cuniv-naama.dz; bou_dje@yahoo.fr; boucetta@cuniv-naama.dz; berrabah@cuniv-naama.dz; hamzaoui@cuniv-naama.dz
Abstract:
The growing volume of Arabic digital content has intensified the demand for intelligent systems capable of processing natural language. Intent classification, a fundamental task in Natural Language Processing (NLP), is essential for enabling effective human-computer interaction by identifying the underlying purpose of user input. However, Arabic intent classification remains challenging due to the language's morphological richness, dialectal diversity, and the scarcity of labeled data. In this paper, we propose a deep learning-based framework for Arabic intent classification, designed to enhance customer support services in social media and e-commerce platforms. A high-quality Arabic dataset was constructed, encompassing three primary intent categories: questions, requests, and complaints. We developed and fine-tuned two deep learning models: a Bidirectional Long Short-Term Memory (BiLSTM) network and a BERT-based transformer model. Both were evaluated using standard metrics, including accuracy and F1-score. Experimental results demonstrate the superior contextual understanding of BERT, while BiLSTM provides a lightweight and efficient alternative for resource-constrained environments. This work contributes to the advancement of Arabic NLP by addressing data scarcity, improving classification performance, and delivering practical tools for real-world applications.
A Hybrid SHAP and Correlation Based Feature Selection Framework for Stroke Prediction
Authors: Noria Bidi (university of mascara)*; Soumia Mohammed djaouti (University of Mascara); Mustapha Sahraoui (University of Mascara)
Emails: n_bidi@univ-mascara.dz; djouti2012@yahoo.fr; sahraoui.musta@univ-mascara.dz
Abstract:
Stroke remains a leading cause of mortality , necessitating the development of effective predictive models for early intervention. This study proposes a robust framework for stroke prediction using the XGBoost algorithm, combining standard clinical features with engineered features, optimized through advanced feature selection techniques. We employ correlation based filtering using Spearman rank correlation to eliminate redundant variables and SHapley Additive exPlanations (SHAP) based ranking to identify the most important features. The selected features are evaluated through a comprehensive set of metrics. Our approach achieved a classification accuracy of 98.1%, F1-score of 98.24%, ROC-AUC of 99.75, and precision of 99.03%. These results advance stroke prediction by simultaneously achieving high performance and deployable efficiency in real-world healthcare settings.
Freeway Traffic State Estimation using EKF-DFNN: Dual Correction Approach
Authors: Asmaa Ouessai (USTO)*; Mokhtar Keche (USTO); Abdekhakim Boudkhil (University of Saida)
Emails: ouessai.as@gmail.com; m_keche@yahoo.fr; boudkhil.abdelhakim@yahoo.fr
Abstract:
One of the significant challenges in modern traffic engineering is the accurate es-timation of the traffic state, this involves precisely determining real-time traffic in-formation, which is crucial for effective traffic management and the implementa-tion of intelligent transportation systems. To address this challenge, various esti-mation algorithms have been developed, most of which are based on the Kalman filter and its variants, offering varying degrees of accuracy. In this paper, we pro-pose an improved traffic state estimation algorithm based on dual state correction using the Extended Kalman Filter and a Deep Feedforward Neural Network (EKF-DFNN). To assess the effectiveness of the proposed algorithm, the esti-mated traffic variables are subsequently utilized as input features for a Support Vector Machine (SVM) classifier, which is employed to predict the traffic state. The proposed algorithm is compared to a baseline approach that combines the standard Extended Kalman Filter with the SVM classifier. The results demon-strate that the proposed EKF-DFNN algorithm outperforms the classical EKF based method in terms of classification accuracy.
Link Quality-based Routing Protocol for Ad Hoc Networks in Urban Environments
Authors: Salah Eddine Benatia (Mascara University)*; Smail Omar (Mascara University); Mekkaoui Abdelkader (Mascara University); Zergaoui Med El Amine (Mascara University)
Emails: se.benatia@univ-mascara.dz; smail@univ-mascara.dz; abdelkader.mekkaoui@univ-mascara.dz; zergaouimedamine@gmail.com
Abstract:
A mobile ad hoc network (MANET) is a set of mobile nodes that create a dynamic topology by cooperating to manage communications. MANETs are characterized by the absence of central administration, reliance on wireless links, node mobility, sensitivity to environmental conditions, and limited energy resources. Routing in such a network is a major challenge due to these constraints. Additionally, environmental factors, such as urban obstacles, can significantly impact link quality and restrict mobility, further affecting routing efficiency. In order to design a reliable routing protocol and resolve this problem, we proposed a multipath routing protocol, LQCA-ue (Routing protocol with Link Quality based for MANETs in Urban environment). We include a fitness function by combining the link quality and stability constraints to select the optimal path. The performance of our proposal is compared with AOMDV and evaluated in four different mobility models, realistic and random. The evaluation of our protocols' results demonstrates that LQCA-UE achieves an 11% improvement in packet delivery ratio and a reduction of up to 50% in routing overhead compared to AOMDV.
Beyond Accuracy: A Rigorous Benchmark and Critical Analysis of Transformer Models for Suicide Ideation Detection
Authors: friki salah eddine (University of mascara)*; Nawaf R. Alharbe (University ofMadinah); Salem Mohammed (University of mascara)
Emails: friki.imed@univ-mascara.dz; nrharbe@taibahu.edu.sa; salem@univ-mascara.dz
Abstract:
Suicide represents a major global public health problem, with hundreds of thousands of deaths each year. Early identification of at-risk individuals is a fundamental prevention strategy, but it is complicated by the significant gap between suicidal ideation and actual suicide attempts. In this context, social media platforms provide an unprecedented source of data for early detection. This paper presents a rigorous benchmark to evaluate the performance of Transformer architectures (BERT) for suicide ideation detection. By comparing a fine-tuned BERT model against a wide range of classical machine learning and deep learning baselines on a large, balanced Reddit dataset, we establish a strong and reproducible baseline for future research. Our best-performing model achieved high effectiveness, indicated by an F1-Score of 0.977 and a ROC AUC of 0.997. However, our contribution extends beyond accuracy; we provide a critical analysis of the methodological limitations, including the implications of using a balanced dataset, and explore model interpretability through self-attention mechanisms. Furthermore, we present a detailed ethical framework for the responsible deployment of these technologies, emphasizing the necessity of human-in-the-loop supervision. We conclude that while advanced language models offer immense potential, their integration must occur within a responsible, human-centered, and clinically validated framework.
Back to Top
