The densification of urban habitats and car-oriented urban planning have led to a growing number of traffic accidents, causing severe casualties, economic losses, and threats to sustainable urban development. Predicting traffic accidents, especially in dense urban areas, is therefore essential for improving public safety and minimizing risks. However, accident sparsity, complex spatiotemporal correlations among diverse risk factors, and external environmental influences make prediction highly challenging. To address these issues, this study proposes a Multi-View Residual Spatio-Temporal Topology Adaptive Graph Convolutional Network (MV-RSTTAG) that integrates heterogeneous risks from road, area, traffic, and external perspectives. By combining topology-adaptive graph convolution, residual connections, and attention-based fusion, the model effectively captures multi-scale spatiotemporal dependencies. Experiments on New York City data (January–June 2016) show that MV-RSTTAG consistently outperforms baseline models across multiple metrics. Efficiency analysis demonstrates millisecond-level inference latency and favorable scalability, supporting real-time deployment. Overall, the model provides an interpretable and efficient framework for intelligent traffic accident risk prediction and urban safety management. • Propose MV-RSTTAG to integrate road and area features for accident prediction. • Introduce deep residual TAG block to capture road network structural info efficiently. • Use attention mechanism and focal loss to enhance feature fusion and handle imbalance. • Analyze TAG’s strengths in capturing complex road network structures effectively. • Show MV-RSTTAG’s superior performance via comparative and ablation experiments.

Eye-tracking desktop activity prediction analyzes how users behave and think through their eye movements for the purpose of behavior prediction during computer use. The study examines how low-point gaze features functioning alongside machine learning (ML) models enable predictions of eight frequent desktop activities which are Debug, Browse, Play, Read, Interpret, Search, Watch and Write. The research uses simple gaze metrics obtained from 24 users through the Tobii X2-30 eye tracker for fixation count analysis along with saccade direction and gaze point statistics in order to support scalable non-intrusive deployment. The research utilized over 200,000 samples from which statistical analytics data was derived along with spatial-temporal eye movement characteristics through preprocessing. Eight well-known ML algorithms including Logistic Regression, Decision Tree and Random Forest, Neural Network and Gradient Boosting, AdaBoost, Naive Bayes and K-Nearest Neighbors received training and evaluation through 80/20 train-test split divisions. Each model conducted testing for activities through the computation of accuracy and precision and recall and F1-score metrics. Results from the evaluations show that Neural Networks coupled with Random Forests produce the best results through average performance metrics which surpass 0.91. The sustained focus activities such as Play, Read, Interpret, Watch and Write responded best to NN while RF demonstrated its strength during tasks of task-switching and problem-solving particularly in Debug and Search activities. The presented study proves the possibility of achieving accurate eye-tracking activity prediction through lightweight gaze features alongside conventional machine learning models. This investigation brings the possibility of creating real-time dynamic interfaces and user-based systems and diagnostic assessments in both consumer and clinical applications. Future research will investigate combining different models with optimization techniques to boost robustness performance in user-variable conditions and dynamic operational settings.

Colorectal cancer is a leading cause of cancer-related mortality; early, accurate histopathological assessment is pivotal. We present a two-stage framework—GR-HDUNET plus a transformer ensemble—that unifies boundary-aware segmentation with multi-model classification. The segmentation stage combines a DenseNet encoder, multi-scale attention, residual decoding, and cGAN refinement to produce anatomically faithful masks, achieving class-average Dice = 0.895, Boundary F1 ≈ 0.82, HOG-similarity ≈ 0.92 on EBHI-Seg. These masks guide a three-branch classifier (Vision Transformer, Swin Transformer, ConvNeXt) whose predictions are fused by weighted averaging. On the six-class EBHI-Seg dataset, weighted averaging attains 95.1% accuracy, 94.2% F1, and micro-AUC = 0.95, surpassing the best single backbone by ≥1.0 percentage point. Qualitative overlays confirm crisp gland delineation, and ROC curves show clear class separation. Importantly, segmentation is trained with pixel-level masks, while classification uses slide-level labels. The pipeline is computationally tractable (≤1.6 s per 224×224 image on a single RTX 4090) and readily reproducible. By coupling cGAN-refined, boundary-aware segmentation with complementary transformer features, our approach delivers robust colorectal-tissue segmentation and classification on EBHI-Seg and offers a transferable template for digital pathology.

The proliferation of hostile content on social media platforms, particularly in low-resource languages such as Hindi, poses significant challenges to maintaining a safe online environment. This study introduces the BGPCN model, which leverages the strengths of Bidirectional Encoder Representations from Transformers (BERT) & Generative Pre-trained Transformer 2 (GPT-2) embeddings integrated with a Relational Graph Convolutional Network (R-GCN) in order to identify hostile information in the language of Hindi. The model addresses both Coarse-grained (Hostile or Non-Hostile) and Fine-grained (Fake, Defamation, Hate, Offensive) classification tasks. The proposed model is evaluated on the Constraint 2021 Hindi dataset, outperforming the latest methodologies in terms of F1-Score of 0.9816, 0.85, 0.50, 0.62, 0.65 regarding both coarse-grained & fine-grained classifications. Comprehensive error analysis and ablation studies underscore the robustness of the BGPCN model while identifying opportunities for refinement. The findings demonstrate that BGPCN offers a reliable and scalable solution for hostile content detection, with potential applications in social media monitoring and content moderation. The data and code will be publicly accessible in https://github.com/mani-design/BGPCN . • Proposed model detects hostile Hindi content using BERT, GPT-2, and R-GCN. • Proposed model excels in coarse and fine-grained classification with high F1-Scores. • Comprehensive analysis shows the proposed model’s robustness and potential for refinement. • Proposed model sets a new standard for hostile content detection in low-resource languages.

Analyzing microbial colonies on agar plates is a critical task in microbiology, with applications spanning research, diagnostics, and industry. Despite advances in automated systems, challenges remain in accurately segmenting and clustering colonies due to variability in their appearance and distribution. To address these challenges, we present a fully automated modular system for colony segmentation and clustering, combining state-of-the-art deep learning for segmentation and machine learning for cluster count prediction and clustering. The main contribution of this work is the proposition, development and evaluation of a novel system, which achieved a V-measure of 0.532 under real-world conditions, improving to 0.727 with ideal segmentation and cluster counts, setting a new benchmark for microbiological analysis. At the core of the system we propose AgarNet, an Attention U-Net-based architecture combined with an EfficientNetB4 backbone, which achieved an F1-score of 0.906 for segmentation. We also introduce a new BRUKERCLUSTER dataset, which is one of the largest and most diverse annotated resources of its kind, featuring expert-annotated images from a controlled cultivation. By combining the dataset, robust segmentation, accurate cluster count prediction, and effective clustering, the proposed system delivers a scalable solution for advancing automated microbial colony analysis. These results establish a new benchmark for automated colony clustering in microbiology, and by releasing the unique BRUKERCLUSTER dataset, we provide a valuable tool for future advancements in automated colony analysis.

FOG computing is an essential paradigm for Internet of Things (IoT) environments, as it enables the decentralisation of data processing to make better use of resources. The rapid proliferation of IoT devices has created new challenges for this middle layer, especially in terms of reconfiguration and dynamic scalability, processes that have traditionally been managed manually. This study presents an architecture for the FOG layer that implements dynamic control and scaling automation through an advanced control layer composed of a coordinator that calculates the optimal distribution of computational load and a balancer that manages traffic routing. The central component of the proposal is the Simple FOG Management Protocol (SFMP), designed specifically to facilitate automated resource management, ensuring efficient allocation of computational capacity, fair load balancing, and robust fault tolerance mechanisms. Experimental validation in several complex IoT environments at the University of Alicante demonstrated a substantial optimisation in resource utilisation with a reduction in energy consumption of between 40 % and 65 %. All this confirms the flexibility and adaptability of the system to dynamic variations in computational demand in large-scale IoT implementations and its benefits over traditional proposals. • Dynamic FOG architecture for load balancing and automatic scalability. • Optimisation of resources and reduction of energy consumption (40–65 %) in the FOG layer. • The SFMP protocol ensures fault tolerance and high availability of the FOG system.