Facilitating citizens with accurate traffic flow prediction increases the quality of life. Roadside sensors and devices are used to capture live streams of huge data and the Internet of Things (IoT) is becoming popular for the deployment of effective Intelligent Transportation Systems (ITS). Traffic flow prediction from the live datastreams require building a data-driven model. This is a challenging task and has attracted researchers for better interpretation of the traffic characteristics. The core problem in traffic prediction is modeling a diversity of traffic trends and unpredictable flow variations with temporal dependencies. Initially, statistical and shallow neural network models were applied to some extent. Recently, deep learning has come up with proven and promising outcomes. Gated Recurrent Unit (GRU) is a variation of recurrent neural networks used effectively for traffic flow prediction. Like other deep networks, GRU uses hyperparameters and a sliding window time-steps mechanism to prepare and tune the model. Better tuning for hyperparameters and search for optimal window size is a tedious process. In this research work, we present an algorithm for hyperparameters tuning along with sliding window steps optimization. Results obtained on a real-time public traffic dataset show a higher capability of the proposed method to reduce the error and an average gain of the optimized model over the untuned network is 4.5%. Furthermore, we apply the optimal hyperparameters obtained in the experiment to other deep learning models and present that our approach improves prediction accuracy and stability.

A Smart City is characterized mainly as an efficient, technologically advanced, green, and socially informed city. An intelligent transportation system (ITS) is a subset area of smart cities that enhances the safety and mobility of road vehicles. It essentially makes travel more convenient, time-efficient and improves the citizens’ quality of life. Accurate and real-time traffic prediction enables law enforcement agencies with well-informed about traffic congestion. However, accurate traffic prediction has been considered a challenging issue. Traffic prediction has restrictions on road network topology and the patterns of dynamic change in time-series data. We propose a novel deep learning framework GCST-GRU, called graph convolutional Spatio-temporal gated recurrent unit, to determine the next traffic state from traffic data. The proposed model learns complex topological structures by capturing a) spatial dependencies from data by using the graph convolution operator, and b) temporal dependencies by using the GRU neural network. Experimental results demonstrate that our framework can obtain complex Spatio-temporal correlations efficiently from the traffic network and perform better than state-of-the-art baseline models on a real-world traffic dataset. The graphical visualization by using convolution operation over the neural network shows that the model outperforms the reachability of the 3-hops neighbor effect in the traffic data graph. Additionally, the training time of the proposed framework is better than the existing state-of-the-art deep learning studies.

Federated learning constitutes a collaborative and shared machine learning paradigm facilitating the joint development of a global model, distinctively addressing privacy concerns while integrating data from various sources. Urban traffic incident prediction (UTIP) tasks inherently require cross-departmental data collaboration, underscoring the significance of federated learning. Specifically, vertical federated learning (VFL) enables multiple participants, each possessing non-overlapping feature subsets, to collectively train predictive models. Recently, researchers have focused on specific VFL issues, such as feature selection and privacy. This study provides a methodology for developing a VFL model utilizing a significant feature selection strategy. The proposed framework is called feature selection-based VFL traffic incident prediction (FSVFL-TIP), and specifically intends to improve incident prediction accuracy. The effectiveness of our suggested model is studied and compared to the baseline VFL model, revealing that our approach outperforms the baseline by 5.7% to 11.6% in test accuracy on two publicly available traffic datasets. Finally, this study explores the improvement in accuracy under various VFL split configurations. The results indicate that VFL is a preferable solution for improved accuracy and communication efficiency while using high-performing feature selection strategies.

This paper proposes an innovative framework for enhancing vehicular edge computing (VEC) efficiency by addressing challenges in traditional intelligent transportation systems and internet of vehicles. Our aim is to improve content prediction accuracy, reduce offload latency, and ensure data privacy by utilizing advanced techniques such as federated learning (FL) and deep reinforcement learning. We propose an in-car popularity-based caching (PBC) asynchronous system, based on FL, for edge computing. Unlike traditional synchronous FL, asynchronous FL allows the global model to be updated without waiting for all cars to complete learning process and upload their individual models, which helps to reduce training time. Integrating VEC with roadside infrastructure system allows real-time data processing by enabling low latency and high-quality services for applications like traffic management and autonomous driving. Through collaborative caching schemes and personalized content delivery, the proposed model optimizes system performance while adapting to dynamic vehicular environments. By prefetching potentially popular content ahead of time and caching it in edge nodes or roadside units, depending on the position and movement direction of the vehicle, the PBC system reduces the delay in content queries. Comprehensive experimental evaluations demonstrate that the PBC technique outperforms existing benchmark caching schemes in reducing the latency.

Large Language Models (LLMs) have been rapidly evolving lately, resulting in the need for strong, explainable models to detect the difference between human-generated and machine-generated articles. Existing approaches which are mostly based on fine-tuned transformers suffer from several drawbacks such as rapid obsolescence, paraphrasing attacks, and lack of interpretability. To improve their ability to detect, this paper proposes a novel paradigm called Human vs. LLM Identification (HLI) which introduces a Retrieval-Augmented Generation (RAG)-inspired evidence-based detection strategy alongside a fine-tuned transformer classifier. Our core model, DeBERTa-Sentinel, is built on top of a fine-tuned Microsoft DeBERTa-v3-small model, which uses a disentangled attention mechanism to better capture subtle syntactic and stylistic deviations characteristic of AI-generated text. We evaluate our framework on a balanced dataset of 43,456 text samples, curated from the OpenGPTText corpus and covering AI-generated and human-authored content across diverse domains including news, education, and creative text. The experimental results show improved performance over the selected baselines, with our framework achieving an accuracy of 97.53%, precision of 95.89%, recall of 99.34%, and ROC-AUC of 99.53%. In addition, explainability is integrated into our framework through Local Interpretable Model-agnostic Explanations (LIME) analysis, providing token-level insight into classification decisions. This study establishes a benchmark for scalable, explainable AI text detection, with implications for academic integrity, content moderation, and combating misinformation.

Breast cancer remains a critical global health challenge with early detection being vital for improving patient outcomes. Traditional diagnostic methods may be time-consuming and resource-intensive, highlighting the need for efficient machine learning solutions. This study addresses this need by developing a robust machine learning framework for breast cancer prediction using the Wisconsin Breast Cancer Diagnostic dataset. We implement a comprehensive preprocessing pipeline, intelligent feature selection, and rigorous comparative evaluation of seven advanced ML models including XGBoost, Neural Networks, and ensemble methods. Our evaluation prioritized both classification accuracy and computational efficiency, explicitly measuring model training and inference time. Results demonstrated exceptional performance with the SGD Classifier achieving the highest test accuracy of $\mathbf{98.25\%}$, while XGBoost, AdaBoost, and SVM RBF Optimized achieved $\mathbf{97.37\%}$ accuracy. The SGD Classifier demonstrated superior computational efficiency, achieving peak performance with a training time of only \textbf{0.05 seconds}, making it significantly faster than other high-performing models. We deployed an interactive Streamlit web application for real-time prediction, bridging the gap between research and clinical practice. This work provides a highly accurate, scalable, and efficient solution for early breast cancer diagnosis, with the code available on our \href{https://github.com/Ibtasam-98/breast-cancer-prediction}{GitHub repository}.