This study examines the performance trade-offs between traditional machine learning methods and large language models (LLMs) in classification tasks across two representative datasets. We evaluate five traditional ML algorithms (Logistic Regression, SVM, Random Forest, XGBoost, and Gradient Boosting) against three LLMs (Llama 3.1 8B, Mistral 7B, and Gemma 3 4B) in both zero-shot and LoRA fine-tuned configurations. Our evaluation focuses on the Activity Recognition using Embedded Mobile Sensors (AReM) dataset for structured sensor data classification and the Twenty Newsgroups dataset for text classification. Traditional ML methods maintain substantial advantages in both accuracy and computational efficiency. On structured data (AReM), traditional ML outperforms the best fine-tuned LLM by 41.4
Accurate classification of knee X-rays across diverse hardware remains a significant challenge due to domain shifts induced by varying scanner specifications, acquisition protocols, and post-processing software. This study investigates out-of-distribution (OOD) generalization techniques to improve the classification of knee calcium deposition across seven distinct X-ray scanners. We evaluated six OOD algorithms, including Empirical Risk Minimization (ERM), SagNet, GroupDRO, Mixup, Adaptive Risk Minimization (ARM), and Invariant Risk Minimization (IRM) to assess the robustness of Vision Transformer-based (DINOv2) versus convolutional (ResNet50) feature extractors within a supervised learning framework. Our experimental results demonstrate that DINOv2-extracted features generalized significantly better to unseen scanners than ResNet50. Among the tested OOD algorithms, SagNet emerged as the most robust approach, achieving 92.6 ± 0.20 and a p-value of 2.9× 10^-10 across the seven scanners by effectively decoupling anatomical content from scanner-specific style. In contrast, IRM exhibited instability across training and testing domains, yielding 85.1
Traditional transportation models inadequately account for how crime risk influences urban mobility decisions, treating safety as a secondary factor despite its significant impact on travel behavior. We develop a multi-task pipeline framework that simultaneously predicts five interconnected outcomes: crime counts, area-level risk classification, route safety assessment, transportation mode choice, and individual vulnerability. Our approach leverages shared feature engineering while maintaining task-specific models to capture complex crime-mobility interdependencies. Using the NetMob Challenge 2025 dataset containing 80,697 trips across 600 French geographic areas with integrated national crime statistics, we employ Random Forest models enhanced with interpretability analysis for actionable insights. The framework demonstrates strong predictive performance across all tasks: crime count regression (R ^2 = 0.847), crime risk classification (accuracy = 90.79
We present an integrated AI framework that predicts song popularity and analyzes sentiment directly from raw audio, using advanced feature extraction and machine learning. The system, implemented as a web application, enables industry professionals to make data-driven decisions based on objective audio indicators. Experiments on real-world data confirm the effectiveness of our approach. We detail the system’s design and discuss future directions for improvement.
Large Language Models (LLMs) have shown significant promise in automating repetitive tasks, particularly in code generation. However, most existing approaches focus on generating isolated code blocks, often lacking global context and coherence across an entire application. Key limitations include the absence of interactive previews, limited support for custom modifications, and insufficient attention to code quality and compliance with specifications. To address these challenges and accelerate application prototyping, we propose a multi-agent approach for generating complete application skeletons from textual specifications. Our framework enables real-time interactive previews and allows users to iteratively refine the generated applications through a conversational agent. To ensure quality and specification compliance, we employ an LLM-based evaluator that assesses the generated code. Once validated, the application skeleton can be downloaded for further development or deployment. This approach aims to bridge the gap between intent and implementation, streamlining early-stage software design and development. We illustrate our approach through a case study, demonstrating the applicability and effectiveness of each step in the proposed method.
Panoramic dental image segmentation is essential for automating diagnostics and treatment planning, yet deep learning models often struggle with limited annotations and anatomical variability, especially under Single-Task Learning (STL) paradigm. This study evaluates the effectiveness of Multi-Task Learning (MTL) versus STL for segmenting panoramic X-rays, using one primary task (individual tooth segmentation) and two auxiliary tasks (binary teeth and quadrant segmentation). With SegFormer-B5 as the backbone, we compare two MTL setups: shared encoder with multiple heads (MTL-MH) and shared encoder with task-specific decoders (MTL-MD). All models are trained and evaluated on the DENTEX dataset under identical conditions. MTL improves performance, increasing tooth Dice from 0.775 (STL) to 0.786 (MTL), quadrant Dice from 0.937 to 0.942, and binary Dice from 0.938 to 0.94. These results confirm that MTL enhances segmentation accuracy and generalization in dental panoramic imaging.
Network Intrusion Detection Systems (NIDS) are critical for safeguarding digital infrastructures against increasingly sophisticated cyber threats. While traditional centralized machine learning models achieve high detection accuracy, they face challenges of privacy, scalability, and data heterogeneity. In this work, we evaluate classical machine learning (Decision Tree, Random Forest), deep learning (DNN), and federated learning (FedDNN) approaches on three benchmark datasets: UNSW-NB15-V2, UNSW-NB15-V3, and NF-BoT-IoT. Our results show that centralized models achieve accuracies above 99.7
Due to the overgrowth of GPS-equipped mobile devices (as smartphones, smartwatches, etc.) and IoT environments, location-based services are generating a large number of objects that contain both spatial and textual information, and therefore, spatio-textual databases have received significant attention in recent years. This paper addresses the problem of Spatio-Textual Similarity Join (STSJ), which outputs object pairs that can be important in applications such as social networking, tourism, etc. This join operation consists of, given a set of objects that contain both spatial and textual information, returning pairs of objects that are spatially close and textually similar. To accomplish this, we design and implement four different algorithms to evaluate the STSJ query using the IR-tree in main memory. Two of them are iterative algorithms (Best-First traversal) and the other two are recursive versions (Depth-First approach). For both iterative and recursive algorithms, we use the plane-sweep technique and the smart all-by-all algorithm to improve query processing. Finally, we conducted extensive experiments on real-world datasets to evaluate the performance of our algorithms for STSJ. The main performance conclusion is that the Best-First algorithm with the smart all-by-all combination has excellent performance for small-medium dataset sizes, while the recursive approach with the same combination reports the best performance for larger ones.
Generative Adversarial Networks (GANs), a prominent subclass of generative models, have gained significant attention in recent years for their ability to replicate complex real-world data. Originating from the concept of a two-player zero-sum game, GANs pit a generator and a discriminator against each other, with the success of one offset by the failure of the other. This framework has driven remarkable advancements across domains such as computer vision and image generation. In this review, we introduce the fundamental concepts of GANs, provide an overview of their major variants, and compare them with other generative models. We discuss the breadth of their applications, analyze different GAN architectures, and situate our findings within the context of prior research.
Decision Support Systems (DSS) have traditionally relied on relational data warehouses and ETL pipelines, which provide stable analytics over structured data but struggle with the rapid emergence of heterogeneous data lakes. Rebuilding DSS infrastructures to natively accommodate NoSQL sources is rarely feasible due to the critical role of legacy systems. This paper addresses the challenge of incrementally extending DSS with evolving NoSQL data while preserving analytical consistency, integrity, and compatibility with existing assets. To this end, we propose an MDE-based approach that leverages a unified model to abstract NoSQL structures into entities, attributes, and relationships. A set of transformation rules is introduced to generate evolution operations that update DSS models and keep them consistent. We demonstrate the applicability of our approach through a medical case study based on the MIMIC-III clinical database, extended with NoSQL sources from telemedicine, Internet of Medical Things, and rehabilitation data. The results show that our method enables seamless model evolution, preserving semantics and unlocking new analytical capabilities without disrupting the operational core.
Online travel reviews contain detailed signals about visitor experience, yet converting this unstructured text into evidence that destination managers and travelers can use is non-trivial. This paper introduces a human-in-the-loop workflow that pairs qualitative coding with lightweight Large Language Model (LLM) analytics. Using Osaka, Japan, as a case study, we compile and pre-process reviews from Google Maps and Tripadvisor, distill them into a 12-topic schema that anchors automated topic and sentiment assignment, and validate model outputs against a human-coded subset using standard metrics. Integrated with exploratory analysis, the results yield a topic–sentiment matrix that highlights strengths and pain points across attractions and services. Findings indicate that a small LLM can deliver scalable, reliable polarity for clearly valenced texts, while the qualitative step sharpens topic granularity and guards against model drift. The pipeline is transparent and reproducible, and it generalizes to other destinations and platforms, enabling evidence-based destination management and more informed trip planning for travelers.
Mobile devices are now indispensable, but their widespread use makes them prime targets for malware. Traditional signature-based methods struggle against sophisticated and zero-day threats, driving the rise of Machine Learning (ML) and Deep Learning (DL) for mobile malware detection. These techniques can generalize from data and adapt to evolving attacks. This survey reviews ML and DL approaches by categorizing studies across analysis strategies (static, dynamic, hybrid), feature extraction methods, and model types (classical ML, DL). It also summarizes commonly used datasets and highlights challenges such as data scarcity and explainability while outlining future research directions. Unlike prior surveys that focus on a single perspective, our work integrates analysis type, feature set, and model to provide a holistic view of mobile malware detection.
Recently, face generation from natural language descriptions has become one of the key applications of generative models and one of the most challenging tasks in computer vision and remains. This text-to-face synthesis technique has a wide range of potential applications, including photo editing, forensic investigation, and game development. While significant progress has been achieved in English-based implementations, extending this technology to the Arabic language presents unique challenges due to the scarcity of Arabic datasets and the linguistic complexity of the Arabic morphology, syntax, and semantics. In this paper, we integrate GAN-CLS, a lightweight yet effective text-to-image generation framework, with the AraBERT, a pre-trained Arabic language model to enable face synthesis from a single-sentence Arabic text description. To support our experiments, we constructed a novel Arabic text-to-face dataset by translating English context descriptions from the Multi-Modal CelebA-HQ dataset into Arabic using DeepL Translator. The performance of the proposed framework was evaluated using the Fréchet Inception Distance (FID) and the Learned Perceptual Image Patch Similarity (LPIPS) metrics. This study constitutes an initial attempt at generating facial images from Arabic textual descriptions.
African cities face unprecedented urban growth challenges, yet lack adequate transport demand models to guide mobility policies effectively. Traditional approaches rely on household travel surveys, which are scarce and outdated in most African contexts. This research explores innovative data sources, particularly camera-based counting data processed through computer vision techniques, to address this critical gap. We present a comprehensive methodology combining mobile phone data, traffic counting from video streams, and machine learning approaches to generate realistic travel demand patterns. Our experimental validation in Libreville, Gabon, demonstrates the feasibility of using commodity smartphones with pre-trained object detection models (YOLOv8) to extract traffic flow data. The results show promising accuracy in vehicle detection and counting, with potential applications for origin-destination matrix estimation and transport planning. This work contributes to the growing field of data-driven urban mobility analysis in developing countries, offering cost-effective alternatives to traditional survey methods while maintaining scientific rigor in transport demand modeling.
Skin diseases pose a growing global health concern that is often underestimated and underprioritized in clinical practice. Timely and accurate diagnosis is vital for effective treatment. Recently, Deep Learning approaches, especially Convolutional Neural Networks (CNNs), have shown promise in automating the diagnostic process. However, as the number of disease classes increases, CNNs and traditional Visual Transformers (ViTs) generally struggle with visually similar diseases, resulting in decreased classification performance. To address these challenges, we propose a new approach that combines two advanced ViT architectures—Shifted Window Transformer (Swin) and Multi-Scale Vision Transformer (MViT)—and ensembles these models using a stacking technique to boost classification accuracy. Using a dataset of 19,559 images containing 23 skin disease classes from a Kaggle directory curated from Dermnet, our ensemble achieves 90
As AI systems are increasingly deployed in high-stakes domains such as finance, healthcare, and education, their responsible use requires scrutiny of how they are engineered, not just how they perform. Although there are now documentation standards for datasets and models, the source code that governs data handling, training, and deployment is rarely audited systematically. To address this gap, this paper introduces Codecard as a part of the System Cards Framework. Codecard is a pipeline that evaluates AI codebases against five criteria: reproducibility, design transparency, documentation quality, privacy, and testing practices. It parses each repository, segments its artifacts, and queries a large language model with structured prompts to produce a scorecard containing numeric ratings, supporting evidence, and targeted recommendations. Codecard’s evaluation of 12 public machine learning repositories revealed that only a minority achieved strong reproducibility and documentation, while testing and modular design were consistently weak. These findings show that code-level audits complement existing dataset and model documentation and can guide concrete engineering improvements.
Internet resilience has become a critical policy concern, as connectivity is increasingly disrupted by both environmental shocks and infrastructural failures. Recent events in Egypt—including severe storms in Alexandria and a major fire at Cairo’s Ramses Central exchange—illustrate these dual vulnerabilities. This paper assesses how meteorological and infrastructure factors jointly shape internet performance across Cairo, Alexandria, and Al Buhayrah. We integrate user-observed throughput and latency from Ookla Speedtest, meteorological reanalysis from ERA5-Land, tower infrastructure from OpenCelliD, and night-time lights from VIIRS into a harmonized, grid-based monthly panel. Ensemble classifiers (Random Forest and XGBoost) were trained on discretized download-speed quartiles, achieving accuracies above 95
Orthopantomograms (OPGs) hold fundamental significance in the practice of diagnostic dentistry. They provide a panoramic image of the maxillofacial region. Deep learning can automate and enhance the analysis of OPGs, which assists clinicians in detecting dental pathologies more efficiently and consistently by utilizing automated systems. In this paper, we propose a solution for multi-class, image-level classification of common dental pathologies using OPG X-rays. A feature extractor with a pretrained ResNet50 architecture was employed, augmented with a Convolutional Block Attention Module (CBAM) to capture significant pathological attention. The dataset was balanced using augmentation methods and contained six classes of dental conditions: BDC-BDR, Fractured Teeth, Impacted Teeth, Caries, Healthy Teeth, and Infection. We present the model architecture, training strategy, and the experimental details for our evaluation. Initial findings indicate the promise of this method, as well as areas for enhancement, such as broadening the approach to multi-label classification and exploring multi-backbone alternatives.
With the increasing competitiveness of today’s job market, job recommendation systems play a crucial role in matching job seekers with suitable opportunities. However, current approaches often struggle to capture the complex relationships between users and job postings due to data sparsity, limited contextual understanding, and difficulties in extracting user preferences from long texts like resumes. To address these limitations, we propose a novel approach that combines a heterogeneous Graph Convolutional Network (HGCN) with Latent Dirichlet Allocation (LDA). While traditional methods fail to model the diverse nature of user preferences and job characteristics, our approach combines the ability of HGCN to capture intricate user-job interactions with the capability of LDA to extract latent topics from job descriptions and resumes. This hybrid approach results in a richer representation of both relational and semantic information, enabling more personalized and accurate job recommendations. Experimental results confirm that our model significantly outperforms existing methods, demonstrating its effectiveness in enhancing recommendation accuracy.
Acne is one of the most prevalent skin disorders globally, especially affecting teenagers and young adults, often resulting in emotional stress and permanent scarring. Although computer-assisted diagnostic tools have been studied, most current approaches lack detection at the individual lesion level and are too computationally intensive for mobile use. This paper introduces an efficient, real-time acne detection system leveraging YOLOv8n, optimized for mobile and edge deployment. The model is trained on a specially labeled dataset with four acne lesion types—papules, pustules, nodules, and comedones—annotated at the lesion level. To overcome data imbalance and scarcity, we utilized image augmentation and per-class replication methods. Hyperparameter optimization was automated using Optuna to further boost model accuracy. The system’s performance was assessed using standard object detection benchmarks, including Precision, Recall, mAP@0.50 (0.501), and mAP@0.50:0.95 (0.162). Results confirm the framework’s capability to deliver accurate lesion-level identification with fast execution speed, supporting its practical use in mobile dermatology applications.
