Khelil Kassoul

High-dimensional similarity search remains a critical challenge in machine learning, particularly when data lie on complex, non-linear manifolds that undermine the effectiveness of classical Locality-Sensitive Hashing (LSH). This work introduces Gaussian LSH, a kernel-based hashing framework that integrates over-clustering with Gaussian probability density modelling to improve locality preservation while maintaining computational efficiency. The method generates compact binary codes from a hybrid kernel–PDF score and supports scalable GPU-accelerated indexing for large datasets. Empirical evaluations across multiple visual and textual benchmarks demonstrate consistent improvements in recall and query latency compared to representative LSH variants and approximate nearest neighbour libraries. Gaussian LSH achieves recall gains of up to $\rm{9}\,\rm{pp}$ and latency reductions of up to $4.3\times$, with benefits sustained across a range of code lengths. These results highlight the approach's scalability and accuracy, supporting its use in medium- to large-scale similarity retrieval tasks across diverse data domains.

This paper presents the Progressive Moving Average Transform (PMAT), a novel signal transformation method for converting time-domain signals into 2D representations by progressively computing Moving Averages (MAs) with varying window sizes. The approach aims to enhance signal analysis and classification, particularly in the context of heartbeat classification. Our approach integrates PMAT with a 2D-Convolutional Neural Network (CNN) model for the classification of ECG heartbeat signals. The 2D-CNN model is employed to extract meaningful features from the transformed 2D representations and classify them efficiently. To assess the effectiveness of our approach, we conducted extensive simulations utilizing three widely-used databases: the MIT-BIH database and the INCART database, chosen to cover a wide range of heartbeats. Our experiments involved classifying more than 6 heartbeat types grouped into three main classes. Results indicate high accuracy and F1-scores, with 99.09% accuracy and 92.13% F1-score for MIT-BIH, and 98.37% accuracy and 79.37% F1-score for INCART. Notably, the method demonstrates robustness when trained on one database and tested on another, achieving accuracy rates exceeding 95% in both cases. Specifically, the method achieves 96% accuracy when trained on MIT-BIH and tested on the ST-T European database. These findings underscore the effectiveness and stability of the proposed approach in accurately classifying heartbeats across different datasets, suggesting its potential for practical implementation in medical diagnostics and healthcare systems.

Tinnitus is the perception of sounds like ringing or buzzing in the ears without any external source, varying in intensity and potentially becoming chronic. This study aims to enhance the understanding and treatment of tinnitus by analyzing a dataset related to tinnitus therapy, focusing on electroencephalography (EEG) signals from patients undergoing treatment. The objectives of the study include applying various preprocessing techniques to ensure data quality, such as noise elimination and standardization of sampling rates, and extracting essential features from EEG signals, including power spectral density and statistical measures. The novelty of this research lies in its innovative approach to representing different channels of EEG signals as new graph network representations without losing any information. This transformation allows for the use of Graph Neural Networks (GNNs), specifically Graph Convolutional Networks (GCNs) combined with Long Short-Term Memory (LSTM) networks, to model intricate relationships and temporal dependencies within the EEG data. This method enables a comprehensive analysis of the complex interactions between EEG channels. The study reports an impressive accuracy rate of 99.41%, demonstrating the potential of this novel approach. By integrating graph representation and deep learning, this research introduces a new methodology for analyzing tinnitus therapy data, aiming to contribute to more effective treatment strategies for tinnitus sufferers.

Epilepsy is characterized by recurring seizures that result from abnormal electrical activity in the brain. These seizures manifest as various symptoms including muscle contractions and loss of consciousness. The challenging task of detecting epileptic seizures involves classifying electroencephalography (EEG) signals into ictal (seizure) and interictal (non-seizure) classes. This classification is crucial because it distinguishes between the states of seizure and seizure-free periods in patients with epilepsy. Our study presents an innovative approach for detecting seizures and neurological diseases using EEG signals by leveraging graph neural networks. This method effectively addresses EEG data processing challenges. We construct a graph representation of EEG signals by extracting features such as frequency-based, statistical-based, and Daubechies wavelet transform features. This graph representation allows for potential differentiation between seizure and non-seizure signals through visual inspection of the extracted features. To enhance seizure detection accuracy, we employ two models: one combining a graph convolutional network (GCN) with long short-term memory (LSTM) and the other combining a GCN with balanced random forest (BRF). Our experimental results reveal that both models significantly improve seizure detection accuracy, surpassing previous methods. Despite simplifying our approach by reducing channels, our research reveals a consistent performance, showing a significant advancement in neurodegenerative disease detection. Our models accurately identify seizures in EEG signals, underscoring the potential of graph neural networks. The streamlined method not only maintains effectiveness with fewer channels but also offers a visually distinguishable approach for discerning seizure classes. This research opens avenues for EEG analysis, emphasizing the impact of graph representations in advancing our understanding of neurodegenerative diseases.

This paper introduces a modeling methodology for Assembly/Disassembly (AD) systems, proposing an effective simulation-based optimization approach. The objective is to simultaneously determine optimal allocations of service times and buffer sizes within unreliable AD systems, ultimately aiming to maximize the production rate. Our method integrates a global search approach, Genetic Algorithm (GA), with a local search method, Finite Perturbation Analysis (FPA). Through this integration, we extract production rate gradient information based on system parameters, which is then incorporated into a stochastic optimization algorithm to simultaneously identify the best allocation of service times and buffer sizes. The proposed approach is applied to various instances of AD systems, and experiments are conducted to assess its performance. The results demonstrate the superior performance of the integrated GA-FPA method when compared to the standalone GA and FPA methods, both in terms of convergence behavior and solution quality. The identified allocation patterns and the potential to expand our approach to other complex manufacturing systems can provide valuable insights for managers seeking to establish more sustainable and efficient assembly and disassembly systems.