http://dspace.univ-setif.dz:8888/jspui/handle/123456789/27 Tue, 12 May 2026 07:10:35 GMT 2026-05-12T07:10:35Z http://dspace.univ-setif.dz:8888/jspui/handle/123456789/6642 Titre: Ethics and Trustworthiness of Algorithmic Decision-Making Systems Auteur(s): Touameur, OUissem Résumé: The rapid adoption of artificial intelligence (AI) in high-stakes domains such as rec- ommender systems, healthcare, and disaster management has amplified the need for trustworthy systems. However, many AI models remain limited by opaque decision- making, biased or noisy data, and the lack of explicit mechanisms to model and quantify trust. This thesis addresses these limitations by integrating trust at three complementary levels—data, model, and prediction—through the combined use of knowledge graphs (KGs) and graph neural networks (GNNs). At the data and model levels, the thesis introduces a taxonomy of trust dimensions, including accuracy, reliability, provenance, fairness, robustness, and explainability, and demonstrates how structured knowledge and graph-based learning enhance transparency and relational reasoning. Building on this foundation, the first major contribution is GUITARES, a trust-aware recommender system based on graph attention networks. GUITARES integrates item confidence derived from external knowledge graphs, inferred user–user trust relationships, and structural learning over user–item graphs. Experimental evaluation shows that GUITARES achieves an RMSE of 0.80, outperforming state-of-the- art baselines while maintaining scalability and robustness. The second major contribution focuses on trust in predictions. The core framework, GraphSkinUQ, is proposed for skin cancer classification, combining CNN feature embed- dings, graph-based relational modeling, and uncertainty quantification to assess predictive confidence. GraphSkinUQ achieves 91% accuracy, with predictive uncertainty between 10% and 11%, a Brier score of 13%, an Expected Calibration Error (ECE) of 6%, and a ROC-AUC of 94%, demonstrating strong performance and well-calibrated confidence esti- mates. This predictive-trust framework is then extended to disaster management through the TDC-GCN model, which adapts the same principles—CNN features, graph convolu- tional learning, and Monte Carlo dropout—to disaster image classification. TDC-GCN achieves 97% accuracy with an entropy-based uncertainty measure of 30%, confirming its effectiveness in high-stakes scenarios. Overall, results across recommendation, medical imaging, and disaster analysis demon- strate that embedding trust mechanisms—from structured data modeling to uncertainty- aware predictions—significantly improves both performance and reliability. This thesis contributes to the development of AI systems that are not only accurate, but also transpar- ent, robust, and trustworthy. Thu, 01 Jan 2026 00:00:00 GMT http://dspace.univ-setif.dz:8888/jspui/handle/123456789/6642 2026-01-01T00:00:00Z http://dspace.univ-setif.dz:8888/jspui/handle/123456789/6626 Titre: Knowledge Discovery in Big Data application to arabic handwriting characters recognition and genomic Auteur(s): Nasri, Khaled Résumé: The exponential growth of Big Data necessitates scalable knowledge discovery methodologies capable of handling massive, heterogeneous datasets. This dissertation explores distributed and parallel deep learning architectures across two critical application domains: genomics and Arabic handwriting recognition. In genomics, a novel distributed pipeline for variant calling integrates multimodal sequencing data through a hybrid Transformer-CNN architecture with attention-based fusion. The system simultaneously processes DNA sequence context via transformer encoders and read alignment evidence through three-dimensional convolutional networks, enabling accurate genomic variant classification while achieving computational efficiency through Distributed Data Parallel (DDP) training across multiple GPUs. Multi-task learning addresses variant type classification, genotype prediction, quality score estimation, and artifact detection, while class-weighted loss functions handle severe data imbalance inherent in genomic datasets. In Arabic handwriting recognition, a hybrid architecture embedding Capsule Networks within Residual Networks (Caps-ResNet) captures hierarchical features and spatial relationships essential for cursive script analysis. This specialized architecture overcomes unique challenges including ligature complexity, diacritical mark sensitivity, and significant stylistic variability across writing styles. This multidisciplinary research demonstrates how hybrid architectures, multimodal fusion mechanisms, and distributed computing strategies enable robust, accurate, and computationally efficient knowledge extraction systems applicable to diverse scientific domains. The results contribute to the development of scalable deep learning solutions that bridge theoretical innovations with real-world applications in genomics and natural language processing. Thu, 01 Jan 2026 00:00:00 GMT http://dspace.univ-setif.dz:8888/jspui/handle/123456789/6626 2026-01-01T00:00:00Z http://dspace.univ-setif.dz:8888/jspui/handle/123456789/6622 Titre: Knowledge Extraction with Machine Learning Techniques from Multi-modal MRI Data application to gliomas classification Auteur(s): Boulkhiout, Youssef Résumé: The present thesis proposes a non-invasive machine learning (ML) framework for predicting MGMT methylation status using features derived from magnetic resonance imaging (MRI) scans, with the ultimate goal of supporting personalized therapeutic strategies. The framework is structured as a three-step pipeline: (i) extraction of imaging features from multimodal MRI; (ii) selection of the most relevant features using Light Gradient Boosting Machine (LightGBM) and Categorical Boosting (CatBoost) algorithms; and (iii) training an ensemble composed of multiple machine learning models on the selected features to classify MGMT methylation status. The model was developed and validated using the Brain Tumor Segmentation (BraTS) dataset, and demonstrated superior accuracy and effectiveness compared to well-known existing approaches. Thu, 01 Jan 2026 00:00:00 GMT http://dspace.univ-setif.dz:8888/jspui/handle/123456789/6622 2026-01-01T00:00:00Z http://dspace.univ-setif.dz:8888/jspui/handle/123456789/6608 Titre: Structured Emotion Analysis from Arabic Text Auteur(s): Senator, Ferial Résumé: Dans le domaine du traitement automatique des langues (TAL), l’analyse des émotions vise à associer un contenu textuel à un ensemble prédéfini d’émotions humaines, incluant généralement la joie, la colère, la peur, la surprise, le dégoût et la tristesse. Les recherches récentes se concentrent principalement sur l’identification des émotions dans les textes en s’appuyant sur des catégories inspirées par les théories psychologiques, telles que les émotions de base proposées par Ekman (1992). Malgré l’importance de la détection des émotions, la majorité des analyses restent superficielles et insuffisantes pour des tâches nécessitant une compréhension plus approfondie du sens émotionnel en contexte. De telles applications exigent de répondre à des questions clés, notamment l’identification de la cause ayant déclenché l’émotion (Cause), la détermination de la personne qui l’a ressentie (Expérient), et, plus généralement, la prise en compte d’informations structurelles telles que qui a fait quoi (Indice), à qui (Cible), pourquoi (Cause) et comment (Manière). Cette thèse doctorale vise à proposer des solutions originales et efficaces pour pallier le manque de ressources et de modèles dédiés à l’analyse structurelle des émotions dans les textes arabes. Pour ce faire, nous introduisons une nouvelle approche d’analyse de la structure argumentaire des émotions en arabe, en tirant parti des avancées récentes des modèles fondés sur les Transformers et, en particulier, des capacités des grands modèles de langue (LLMs) pour l’arabe. Les principales contributions de cette thèse sont multiples. La première contribution consiste en la construction et l’annotation du premier corpus arabe dédié à l’analyse structurée des émotions, nommé ”AraERL”. La thèse propose également une étude approfondie de l’impact de chaque argument sémantique sur la performance de l’identification des émotions. Elle explore ensuite l’utilisation de ChatGPT pour annoter des textes arabes avec des rôles sémantiques et des émotions à travers une approche de projection interlinguale. Ce travail évalue également la capacité de ChatGPT à projeter avec précision en arabe les annotations sémantiques et émotionnelles issues de l’anglais. Enfin, il offre une comparaison complète des performances des modèles ouverts de grande taille (open-LLMs) pour ces différentes tâches. Wed, 01 Jan 2025 00:00:00 GMT http://dspace.univ-setif.dz:8888/jspui/handle/123456789/6608 2025-01-01T00:00:00Z