Statistics, Optimization & Information Computing

Preface for the Special Issue of the fifth International Conference of Computer Science and Artificial Intelligence’24, ICCSAI'24

Oumayma BANOUAR — 2026-01-19

The main contributions of [Stat. Optim. Inf. Comput. Vol.15, No.2 (2026): ICCSAI'24], consisting of fifteen papers selected and revised from the fifth International Conference of Computer Science and Artificial Intelligence’24, are highlighted.

Deep Learning-Based Classification of Retinal Pathologies

Kawtar NAIM — 2025-10-15

Age-related Macular Degeneration (AMD), Diabetic Macular Edema (DME), Epiretinal Membrane (ERM), Retinal Artery Occlusion (RAO), Retinal Vein Occlusion (RVO), and Vitreomacular Interface Disease (VID) are prevalent eye conditions that can lead to partial or complete vision impairment and blindness. Addressing these challenges in eye care necessitates advanced imaging technologies like Optical Coherence Tomography (OCT). The evolution of OCT from time-domain to frequency-domain techniques has significantly enhanced its utility in routine clinical procedures. This paper introduces a novel R50-CapsNet architecture designed to classify retinal diseases more accurately and reliably. Our approach aims to improve diagnostic accuracy for the OCTDL and Kermany datasets.

An Efficient Deep Learning Approach based on 3D Res-UNet for Multimodal Brain Tumor Segmentation

Khaoula Echine — 2025-10-15

Accurate segmentation of brain tumors in MRI is an important aspect of accurate clinical diagnostics and sound surgical planning. Tumor boundary accuracy is fundamental to informing assessment of a patient’s condition, acquired through continuous expansion of Deep Learning based segmentation approaches. This study introduced an effective deep learning method to perform 3D segmentation of multimodal MRI images by enhancing the Res-UNet architecture. Our proposed model, 3D ASPP-ResUNet, incorporates an ASPP (Atrous Spatial Pyramid Pooling) module to better exploit multi-spatial scale features. The BraTS 2020 dataset has been used for training and evaluation. This model performs well according to the dice metric for different tumor regions attaining Dice scores of 0.7442 for TC (tumor core), 0.7293 for ET (enhancing tumor) and 0.8215 for WT (whole tumor). Furthermore, we observed that the 3D ASPP-ResUNet was better than currently leading models with respect to segmentation performance metrics that we defined as the Dice coefficients.

Analyzing Performance Discrepancies: U-Net vs TransUNet for Aircraft Emergency Landing Site Detection

Adil Illi — 2025-10-15

In the context of aviation, forced landings are unwanted events that can happen to an aircraft during its flight trajectory. They can be due engine malfunctions, adverse weather conditions and other sudden situations. For this reason, and to ensure passengers' safety, it is imperative to develop methods and procedures to detect potential sites that can be used as emergency landing areas during these crisis situations. Traditionally, pilots use visual indicators to detect such landing sites, this ability can varry from a pilot to another depending on experience, aircraft altitude and other environmental conditions. Such circumstances can make this visual detection task highly difficult.

Image segmentation is one of the possible solutions that can be implemented in identifying potential emergency landing sites for aircraft. Precise segmentation should improve on the effective identification of safe landing areas, thereby enhancing aviation safety protocols in general.

In this context, the traditional U-Net \cite{unet} architecture has shown exceptional results regarding segmentation tasks. However, a new approach derived from U-Net and incorporating transformers in its encoder, known as TransUNet, has demonstrated promising results, surpassing in some cases those of U-Net.

This study investigates the performance of TransUNet compared to traditional U-Net for aircraft emergency landing site detection. Both architectures were implemented, trained, and evaluated using our novel dataset tailored for this purpose. Our work demonstrate that U-Net outperforms TransUNet in terms of accuracy and computational efficiency in this specific segmentation task. In particular, U-Net exhibited superior performance by improving segmentation precision from 80% up to 88% in the testing set. Moreover, the mean Intersection-Over-Union, a metric for segmentation accuracy, have also seen an improvement of 77% for U-Net over 73% for TransUNet. These results emphasise the power of the traditional U-Net architecture for this critical application, underlying its practical relevance in enhancing aviation safety.

Attention-Enhanced Deep Convolutional Denoising Autoencoder for Cervical Cancer Image Quality Improvement

Salma OUSSAHI — 2025-10-16

Cervical cancer is the main cause of death among women worldwide, and catching it early can make all the difference. Pap-smear images are a cornerstone of screening, but in practice they’re often affected by noise tiny specks, uneven lighting or blur that can throw off both human experts and automated algorithms. To address this, we’ve built an Attention-Enhanced Deep Convolutional Denoising Autoencoder (AE-DCDA). By weaving an attention mechanism into a classic encoder decoder structure, our model learns to focus on the important cell structures and suppress the noise around them, preserving the fine details that matter for diagnosis. We tested AE-DCDA on the Herlev dataset, which originally includes 917 cervical cell images spanning seven classes. To give the model more varied samples, we applied a range of image augmentation techniques such as rotations, flips and small shifts, effectively enlarging our training pool. When we evaluate noisy images held back, our denoiser pushed the Peak Signal to Noise Ratio up to 35.2 dB and drove the Mean Squared Error down to 0.0003, notable gains over conventional filters. In practice, that means clearer cell boundaries, crisper nuclei and fewer artifacts, paving the way for more reliable downstream segmentation or classification.

Innovative Hybrid Techniques for Cloud Detection and Segmentation Using Computer Vision and Machine Learning

Laila AMIR — 2025-11-02

Cloud detection and segmentation play a critical role in satellite imagery analysis and environmental monitoring. This paper presents a novel hybrid approach that integrates traditional computer vision techniques with advanced machine learning algorithms to enhance both accuracy and efficiency in cloud detection systems. The hybrid methods incorporate image processing techniques such as HSV thresholding, morphological operations, histogram equalization, and Canny edge detection, alongside ensemble learning models like Random Forest, SVM, K-Means clustering, and XGBoost. These hybrid approaches outperform standard methods both in terms of accuracy and computational efficiency, with some hybrid methods offering up to 15% higher accuracy and 70% faster processing times compared to their standard counterparts. These findings highlight the potential of hybrid techniques to significantly improve real-time cloud detection performance.

An Evaluation of Discretization Techniques for HMM-Based Classifiers

Boutaina ouriarhli — 2025-11-07

Discretization of continuous features is an important task to handle problems with real values in machine learning. Many supervised classification algorithms perform well using a discrete space and the discretization process of the continuous features is suitable for more traditional algorithms the process. In this paper, we present the classification model based on the Hidden Markov Model (HMM) developed recently by Benyacoub and al using several discretization methods existing in the literature to construct the classifier. We conduct an experimental study using 9 benchmarking data sets to evaluate the performance and examine the effect of discretization methods on the assessment of the proposed learning algorithm. {We conduct an experimental study using 9 benchmarking data sets to evaluate the performance and examine the effect of discretization methods on the assessment of the proposed learning algorithm. We report Accuracy (ACC) and Area Under the Curve (AUC), and we validate the global and pairwise differences across methods using the Friedman test followed by the Nemenyi post-hoc procedure.

A Fisher–Yates Shuffle in a Hardened Merkle–Damgård hash for the blockchain's PoW

Asmaa CHERKAOUI — 2025-11-08

In this paper, we introduce a Fisher–Yates shuffle for the development of the Merkle-Damgård construction while not using any predefined functions or hashing library. Since SHA-1 has been deprecated, we focus on the Secure Hash Algorithm 2 (SHA-2), which remains secure against all known full-round collision attacks. In this work, we introduce and study Fisher--Yates–driven dynamic permutations within this family to enhance resistance against automated cryptanalysis, particularly SAT-based attacks, while preserving SHA-2’s robust design. Finally, we provide a practical explanation of how the use of our approach could be beneficial for the Proof-of-Work (PoW) in blockchain.

An Extended Grendel Approach Applied to Blockchain Signature as an Alternative to Keccak Permutation

ABDELKARIM LKOAIZA — 2025-11-08

In this paper, we present our own developed programming which helps to generate a sponge-based function while avoiding any call from hashing libraries. Then, we try to implement it in a blockchain signature by getting inspired from Keccak methods such as the recently inextinguishable Secure Hash Algorithm 3 (SHA-3), but before this, we note that our main contribution here, is about introducing the Grendel permutation instead of the Keccak one as they both rely on sponge-based procedures, but the shuffling step is different. In fact, even our Legendre symbol considered here, extends the Euler criterion that is restricted to prime field, to the cases of the group of invertible elements Z/pqZ. To the best of our knowledge, this is the first time that such an approach is used in blockchain signature.

Applying Author Profiling On Reddit Comments At The Document-Level

Idriss Oulahbib — 2025-11-24

Author Profiling (AP) encompasses the task of discerning an author’s biological, psychological, and socio- cultural attributes, including but not limited to gender, age, religion, profession, and personality, from their written content. This task is commonly approached as a form of text classification, where models are trained using features extracted from the author’s text to predict labels such as gender and age category. This study investigates the effectiveness of Machine Learning (ML), Deep Learning (DL), and Transformer-based models for age and gender classification at the document level on a large dataset of Reddit comments annotated using Regular Expressions (REGEX). We employed various algorithms, including Naive Bayes (NB), Random Forest (RF), Logistic Regression (LR), Multi Layer Perceptrons (MLP), Convolutional Neural Networks 1 Dimension (CNN1D), and Distilled Bidirectional Encoder Representations from Transformers (DistilBERT). For feature extraction, we utilized Bag Of Words (BOW), Term-Frequency Inverse Document Frequency (TF-IDF), dictionary scores from Linguistic Inquiry Word Count (LIWC), averaged FastText embeddings (both pre-trained and trained on Reddit), and concatenated Subreddit embeddings to enhance contextual representation. Our experimental results revealed that traditional ML models with TF-IDF features, particularly LR, achieved competitive performance compared to deeper architectures. The best accuracy for gender classification was obtained by the DistilBERT + Subreddit embeddings model with 0.65 at the document level and 0.80 at the author level using majority voting. For age classification, the highest accuracy reached 0.37 with the same model configuration, outperforming all baseline approaches. These findings demonstrate that Transformer-based models enriched with contextual features offer a significant improvement over ML and traditional DL models in document-level AP.

Resolution of linear interval systems using neural networks and their application to the Leontief economic model

Mohamed Amine Benhari — 2025-12-02

Linear systems with interval coefficients are a class of mathematical modeling problems whose coefficients do not have exact values, but are multi-valued sets of possible values. This feature characterizes many processes in the real world, especially in economics. The solution of these systems is essential for making decisions under uncertainty. In machine learning, neural networks represent an excellent tool for solving this problem. The performance of neural networks is largely due to their flexibility; they can learn multifaceted dependencies without any prior assumptions about the distribution of input data. In addition, it is important when the original data are measured inaccurately and/or noised. Such data are common in economic forecasting. In addition, their ability to learn data structures enables them to make accurate forecasts based on new data, which is decisive for risk management and strategic decision-making.
One possibility is to model a concrete application on the Leontief model, which describes the flow between individual sectors of the economy. This way, when neural networks are integrated into linear systems with interval coefficients, it is possible to forecast the impact of variations in demand and supply over the whole economy, reducing the uncertainties of economic activity forecasts. In conclusion, the new methodology of neural networks in the resolution of linear systems with interval coefficients may, at some point, prove essential progress in managing economic uncertainty, allowing companies and decision-makers to navigate more confidently in a complex and unpredictable environment.

IoT-CR: A Novel IoT-Based Approach to Address Data Sparsity in Context-Aware Recommendation Systems

Mohamed El Amine Chafiki — 2025-12-10

In recent years, recommendation systems (RS) have become an essential part of modern online services, helping users discover new products, content, and experiences that match their preferences and needs. In particular, context-aware recommendation systems (CARS) have received considerable attention because of their capacity to use contextual information to deliver relevant and personalized recommendations, compared to traditional recommendation systems that only use user preferences and item attributes to make recommendations. However, the performance and effectiveness of CARS are challenged by the rise of data sparsity, a common issue in many recommender systems. It occurs when there is an insufficient amount of user-item interactions. This study explores using varied IoT contextual data to address this issue. We present and assess an IoT-assisted Contextual Recommendation (IoT-CR) system, which is an end-to-end deep learning framework architecture that aims to incorporate rich contexts from IoT sensors seamlessly into the recommendation process. To prove this concept, we perform an extensive comparative study against a set of baseline models on four different public, context-rich datasets. We find a mixed bag of results, indicating that the performance of models depends very much on the characteristics of the datasets such as their size and sparsity. In particular, the IoT-CR framework achieves the best results on the largest dataset where there is enough data for it to learn complex interactions. On the other hand, in smaller or more sparse situations, classical collaborative filtering or tree-based models perform better. This research offers an important benchmark, stating that although supplementing data with IoT signals is a very good way forward, the effectiveness of complex models is not a general case and depends critically on the data landscape.

Enhancing Recommender Systems through Active Learning Strategies with Matrix Factorization

Bachir Asri — 2025-12-11

With the rise of information overload and a multitude of options, the significance of recommender systems as indispensable tools remains paramount. These systems offer personalized suggestions, greatly improving user experiences in navigating the vast array of available options. This study explores the application of Matrix Factorization combined with active learning techniques to enhance the accuracy of recommender systems and tackle frequent challenges like sparse data and the cold start issue. The active learning strategies employed encompass both personalized approaches, like similarity to profile, highest predicted, and binary predicted, as well as non-personalized methods including random, popularity, Gini, error, popgini, and poperror. By applying these strategies to Matrix Factorization using the MovieLens and GoodBooks datasets, the study demonstrates significant improvements over conventional approaches, highlighting the critical role of active learning in enhancing recommender systems to better capture diverse user preferences.

A Contextual Wellness Recommender System

Aicha AIT ELBAZ — 2026-01-04

Since their introduction, personalized recommendation systems have experienced remarkable evolution, aiming
to provide recommendations corresponding to the needs and preferences of users, particularly in the field of healthcare. However, these systems have encountered limitations, particularly regarding the dynamic nature of users health needs. Contextual recommender systems take this dynamic into account and use users contextual data to generate personalized recommendations tailored to their needs. In this context ,Internet of Things (IoT) technologies assume a pivotal role, enabling remote monitoring of various health aspects, facilitating proactive healthcare interventions, and crafting personalized treatment plans. The data collected by IoT devices serves as a valuable resource, enhancing the effectiveness of personalized recommendations.To address this challenge, we propose a novel approach named “Contextual Wellness Recommender System” ,a methodology that fully exploits the data collected by several connected health devices. Our methodology relies on the use of advanced machine learning and data analysis techniques to intelligently integrate contextual information into the recommendation process.Using machine learning algorithms, we will train two models to recognize patterns and correlations in IoT sensor data and other contextual factors. The used data includes users physical activity,demographics, lifestyle habits and vital signs. By taking into account these multiple dimensions of data, our models will be able to generate personalized recommendations that will allow users to proactively take care of their health.With an accuracy of over 90\% on Model\_A and more than 80\% on Model\_B on both training and validation data,our proposed approach stands out for its use of several and diverse connected health devices to generate recommendations. This innovative approach increases efficiency, personalization and adaptability by adapting to different individuals health conditions and fully exploiting their contextual data.

Comprehensive Study: Machine Learning and Deep Learning Approaches in Intrusion Detection Systems

Najoua Azizi — 2026-01-04

This paper presents a synthesis of approaches from various studies aimed at enhancing attack classification using machine learning (ML) and deep learning (DL) models. The works studied cover diverse aspects of cybersecurity, with a particular focus on intrusion detection systems (IDS) and Internet of Things (IoT) security. The paper provides an overview of the datasets used to train ML and DL models, the metrics used to evaluate the performance of these techniques, outlines the process for implementing them, and discusses perspectives and future research directions.

A Hybrid Multi-Objective Immune Algorithm for Commercial Recommendation Systems

Fatima Ezzahra ZAIZI — 2026-01-12