Coronary Artery Disease (CAD) remains one of the leading causes of mortality worldwide and accounts for a great number of prolonged hospital stays, which increase the cost of healthcare and its resource consumption. This research presents Graph Neuro Boost, a new method of predicting hospital LOS for CAD patients based on hybrid feature selection and classification, and characterized by high accuracy of prediction. Using the MIMIC-III dataset, the method involves complete data preparation, analysis, and preprocessing. A Graph Neuro Feature Selector was built by combining the autoencoder-based embedded method and the filter method using graph theory and Kruskal’s algorithm. The method efficiently identified 16 critical features in just 2.68 seconds, surpassing traditional hybrid feature selection approaches. Those features were then optimized using Bayesian optimization together with XGBoost, implementing a balanced Expected Improvement and Upper Confidence Bound technique to reduce log loss. Graph Neuro Boost resulted in an accuracy of 98.57%. In addition to superior accuracy, Graph Neuro Boost demonstrated outstanding performance relative to baseline models on multiple evaluation metrics. The SHAP analysis was also employed to interpret feature contributions, improving model transparency and supporting clinical decision-making, patient management, and efficient healthcare resource allocation.
As technology advances, analyzing Global Positioning System (GPS) data is critical to improving location accuracy in product distribution. In this study, the geolocation accuracy of a distribution company in Istanbul was evaluated using 4,784,003 GPS records and prediction models, including machine learning, deep learning, ensemble and hybrid approaches. The models were tested with both one-dimensional (1D) and four-dimensional (4D) inputs to assess differences in performance. The eXtreme Gradient Boosting (XGB) model achieved the highest accuracy, and its performance improved significantly with SpatioTemporal (ST) feature augmentation. Novel attention-based models, including spatial attention and ST transformer, were proposed to better capture spatial and temporal dependencies. Federated Learning (FL) was integrated into the ST transformer to enable decentralised training and improve privacy. To test the generalisation of the model, the experiments were repeated with the PortugalGPS dataset. The results show that ST features significantly improve positioning accuracy and that machine learning models—especially XGB—outperform deep learning models on urban GPS prediction tasks. This study provides a scalable, privacy-preserving framework for improving geolocation accuracy in different cities and conditions. It also contributes to a better understanding of urban logistics by highlighting the role of GPS data in optimising location accuracy during distribution.
This study introduces a computational framework for analyzing the tonal structure and interactional dynamics of political discussion programs broadcast on Turkish television. Using a self-constructed dataset of televised debates featuring journalists, politicians, and commentators, we trained a CNN on MFCCs to classify speech segments into three tonal categories: sociable, discuss, and monotone. The model achieved strong performance, with an average classification accuracy of 92.3% across five fold cross-validation. To assess the role of individual speakers, we modeled participant influence using Ridge Regression with a CLR transformation, estimating how each participant contributed to the overall tonal composition of episodes. In addition, the study introduces several structural analyses of tonal behavior, including tonal trajectory analysis to observe how tone evolves throughout episodes and segment-to-segment transition modeling to capture sequential tone shifts. Furthermore, entropy-based diversity metrics are employed to assess tonal balance and variety across episodes. Together, these components offer a multi-dimensional view of televised political discourse, revealing both who shapes the tone and how tonal dynamics unfold over time. The findings contribute to computational media studies, political communication, and the analysis of broadcast interaction in polarized environments.
Electroencephalogram (EEG)-based brain–computer interfaces (BCIs) decode neural intent but remain vulnerable to non-stationarity, low signal-to-noise ratios, complex spatiotemporal dynamics, and inter-trial variability. Existing deterministic deep models often fail to capture global temporal dependencies while preserving local spatial topology, limiting robustness and generalization. We propose the Variational Spatiotemporal Enhanced Transformer (VSTE-Transformer), an encoder–decoder framework for robust EEG recognition. It embeds variational inference into a Transformer, representing latent features as distributions to quantify uncertainty and reduce sensitivity to noise. A Spatiotemporal Enhancement Module combines localized graph attention, which models dynamic electrode correlations, with multi-scale dilated convolutions that capture rhythmic temporal contexts. A dual-task objective jointly optimizes classification and signal reconstruction, encouraging the latent manifold to preserve neurophysiologically meaningful structure. Experiments on BCI Competition IV-2a and OpenBMI under a unified within-subject, cross-session protocol show that VSTE-Transformer significantly outperforms state-of-the-art methods and remains more robust under severe noise corruption.
Modern cybersecurity infrastructure management is facing the challenge of an explosion in the number of machines and Key Performance Indicators(KPIs). As a fundamental data-mining technique, clustering can not only group data with similar characteristics but also provide informative insights that are critical for cybersecurity operations. In particular, the clustering of large-scale KPIs can significantly assist cybersecurity engineers in detecting system failures, reducing redundant alarms, and decreasing troubleshooting overhead-contributing to more effective threat detection and alert management. However, KPIs are reasonably long (that is, high dimensionality) because of the continuous collection of the monitoring system. For KPIs’ high dimensionality, existing KPIs clustering methods always have a trade-off between clustering quality and computational complexity. In this paper, we propose a novel method, SubCluster, which is a fast, accurate, and scalable approach for clustering large-scale KPIs with high quality. SubCluster aims to extract representative subsequences from numerous KPIs. In contrast to existing subsequence extraction-based clustering methods that compress the candidate set, SubCluster first conducts coarse-grained clustering with the nearest neighbor-based graph to obtain near-accurate groups from raw KPIs. SubCluster then attempts to extract representative subsequences for fine-grained clustering. Extensive experimental results on public time-series and large-scale KPIs datasets from the real world demonstrate the effectiveness and efficiency of SubCluster.
Children's attention monitoring is crucial in education for improving learning outcomes. However, existing methods often rely on single-modality inputs, making it difficult to address the complexity of children's emotional expressions and attention fluctuations. This paper presents CSW-FocusNet, a multi-modal affective computing framework that integrates visual, auditory, motion, and physiological modalities, combined with temporal modeling and personalized feedback mechanisms for real-time monitoring of children's attention. CSW-FocusNet consists of three key modules: the CENet module for multi-modal feature fusion and dynamic relation capture, the STFEformer module for temporal feature modeling and attention prediction, and the SWO module for personalized feedback strategies. Extensive experiments on the different datasets demonstrate that CSW-FocusNet outperforms 9 baseline models across all metrics. On the DEAP dataset, its Macro-F1 is 0.85, MCS is 0.89, and RMSE is 1.32, significantly exceeding other methods. Future work will focus on refining feedback strategies for improved adaptability and real-time performance.
Convolutional Neural Networks (CNNs) surpass human-level performance on visual object recognition, yet their behavior differs from humans in important ways. One prominent example is that CNNs trained on ImageNet exhibit a texture bias, while humans show a strong shape bias. Although CNN shape bias can be increased through e.g., data augmentation or additional training, the source of this discrepancy remains unclear. Developmental research suggests that one factor driving human shape bias is that during early childhood, toddlers tend to fill their field-of-view with close-up objects. We operationalize this close-up as a zoom-in on objects during CNN training, which increases shape bias without additional training or data augmentation. Systematic manipulation of background-object ratios during training reveals a strong inverse correlation with shape bias. Notably, zooming in on objects, thereby more closely emulating child vision, aligns classification accuracy and shape bias between humans and CNNs. Finally, we achieve a near human-like shape bias when using a developmentally-inspired background-object ratio for training and shape bias assessment. These findings demonstrate that a simple adjustment to image datasets — zooming in on objects — can produce human-like shape bias. This suggests that human learning strategies offer a promising avenue for developing human-aligned, efficient, and robust vision CNNs.
Image restoration is a fundamental problem in computer vision, aiming to recover high-fidelity visual content from degraded observations. Although diffusion-based models have achieved remarkable success, most existing methods operate without explicit high-level semantic reasoning, often producing restorations that lack contextual consistency. To overcome this limitation, we introduce Semantic Refinement Diffusion (SR-Diff), a novel diffusion-based image restoration framework that incorporates semantic priors derived from vision-language models (VLMs). Specifically, a pre-trained VLM is used to extract semantic embeddings directly from degraded inputs. These embeddings are then purified through a Semantic Refinement Module, which removes degradation-induced distortions while preserving semantic coherence. The refined embeddings serve as explicit semantic guidance for the diffusion denoising process via a lightweight cross-attention pathway. Importantly, our semantic guidance introduces minimal inference overhead—requiring only a single VLM forward pass and negligible additional computation during sampling—thus ensuring practical deployment while significantly improving restoration quality. Extensive experiments on diverse image restoration tasks, including dehazing, deraining, motion deblurring, and low-light enhancement, demonstrate that SR-Diff consistently outperforms state-of-the-art approaches in both quantitative metrics and perceptual realism. Ablation analyses further confirm that semantic guidance substantially enhances restoration fidelity, especially in complex scenes requiring contextual understanding. Code, dataset and models are available at: https://github.com/ZhuQ1ang/SR-Diff
In the Internet of Medical Things (IoMT), reliable electroencephalogram (EEG) analysis is crucial for early diagnosis of schizophrenia and other neurological disorders. However, limited schizophrenia EEG data restrict the generalization ability of deep learning models, hindering clinical deployment. Existing augmentation methods often produce low-quality synthetic samples or suffer from overfitting and vanishing gradients. To address these issues, we propose IGANEEG, an improved generative adversarial network for small-sample schizophrenia EEG augmentation. IGANEEG incorporates joint time-frequency-spatial feature groups from real EEG as conditional inputs to generate realistic synthetic samples retaining critical data characteristics. The generator is replaced by an autoencoder to reduce reconstruction loss, and a dynamic pause mechanism is proposed to enhance training stability while the learning capacity of the hidden layers is optimized. Experiments on a public dataset and a private dataset show that IGANEEG achieves recognition accuracies of 96.8% and 98.2%, respectively, substantially outperforming state-of-the-art methods. These results validate the framework's effectiveness in improving model generalization and offer valuable data support for intelligent healthcare systems in the IoMT.
Accurate radio map construction is essential for 6 G wireless network optimization, yet faces significant challenges due to sparse real-world measurements and dynamic environmental obstacles. This paper presents RMF, a novel one-step generative model based on mean flow matching that enables direct mapping from noise to radio map distribution in a single forward pass. Our approach integrates a multi-feature U-Net architecture with specialized branches for processing building layouts, base station configurations, sparse measurements, and dynamic obstacles through cross-attention fusion. Extensive evaluations on the RadioMapSeer dataset demonstrate that RMF achieves state-of-the-art performance, reducing RMSE by 7.5-12.2% compared to diffusion-based methods while maintaining competitive SSIM scores of 0.9557-0.9674. In challenging zero-measurement scenarios, RMF attains PSNR improvements of 1.45-1.65 dB over existing approaches, showcasing robust performance in both static and dynamic environments. The model's balance of accuracy and efficiency makes it particularly suitable for real-time 6 G applications including coverage optimization and dynamic resource management.
Efficient emergency vehicle routing in urban environments is critical for timely medical response, yet existing approaches often decouple traffic prediction from routing, overlook heterogeneous uncertainty, and lack causal reasoning under routing interventions. We propose an end-to-end differentiable framework that integrates risk-aware routing, causal traffic forecasting, and decomposed uncertainty quantification. Specifically, a regime-conditioned evidential heterogeneous spatiotemporal graph neural network models traffic dynamics on heterogeneous road networks while separately estimating aleatoric and epistemic uncertainty across road segments. To capture intervention effects, a causal graph neural network learns dynamic causal dependencies and enables counterfactual prediction of traffic changes induced by emergency routing decisions. Building on these components, we design a multi-objective routing strategy that adaptively balances travel time, reliability, and safety risk according to real-time hospital capacity and patient injury severity. Experiments on METR-LA and PEMS-BAY demonstrate improved prediction accuracy, better-calibrated uncertainty estimates, and more clinically informed adaptive routing than strong baselines. The proposed framework provides a practical and reliable solution for safety-critical emergency transportation in complex urban traffic systems.
Vehicle tracking in complex traffic environments remains a challenging task due to frequent target occlusion, high visual similarity among vehicles, and cluttered backgrounds. These factors often result in tracking drift and target misidentification, undermining the reliability of tracking systems in real-world applications. To address these challenges, this paper presents L2G-TransT, a vehicle tracking framework based on local-to-global multi-scale attention. The proposed method hierarchically integrates fine-grained local features, mid-scale spatiotemporal context, and global semantic cues through a combination of window attention, shifted window attention, and global self-attention mechanisms. A dynamic weighting strategy is employed to adaptively fuse multi-scale features based on the target’s visibility and appearance changes, thereby enhancing the model’s adaptability to challenging scenarios. In addition, a multi-head cross-attention module is introduced to improve feature discriminability by suppressing background noise and emphasizing target-relevant regions. Extensive experiments demonstrate that L2G-TransT achieves better tracking performance, improving both accuracy and robustness under occlusion and appearance ambiguity. The codes and datasets are available at https://github.com/chd-via-lab/L2G-TransT.git.
In the field of recommendation systems, click-through rate (CTR) prediction is essential for measuring user engagement and product interest. However, input features are typically high-dimensional and sparse, requiring effective modeling of high-order feature interactions. Existing methods learn low-dimensional representations and identify useful feature combinations, but they often struggle to capture implicit interactions in non-Euclidean spaces, leading to noisy feature relationships, reduced interpretability, and suboptimal prediction performance. To address these challenges, we introduce a similarity-based adaptive graph attention network (SGAT) for CTR prediction. The SGAT employs a dual-attention mechanism that integrates similarity-based adaptive attention with softmax attention to selectively emphasize beneficial feature interactions while suppressing irrelevant ones. This strategy reduces noise, enhances interaction modelling, and mitigates overfitting. Experimental results on two public benchmark datasets show that SGAT achieves superior performance compared with several state-of-the-art baseline methods.
Diabetic retinopathy and glaucoma are significant causes of blindness worldwide, affecting one-third of diabetic patients and indicating an increased risk of vascular complications. This study proposes a modified DeepLabV3+ semantic segmentation model for segmentation and a deep convolutional neural network (DCNN) model to classify and grade diabetic retinopathy and glaucoma, namely, the modified DeepLabV3+ segmentation and classification model (MDSCM). This study utilises two proposed datasets, namely, KEH_3815 and KEH_4129, which contain 8000 images collected from Khulna Eye Hospital and Laser Centre, Khulna, Bangladesh, and eight other benchmark datasets to evaluate the proposed methods. The proposed modified DeepLabV3+ semantic segmentation model achieves 3.69%, 2.92% and 5.38% higher accuracy, F1-score and Jaccard index than the original DeepLabV3+ semantic segmentation model in diabetic retinopathy segmentation. Similarly, for glaucoma, the modified DeepLabV3+ semantic segmentation model achieves 3.07%, 3.19% and 6.11% higher accuracy, F1-score and Jaccard index, respectively, than the original DeepLabV3+ semantic segmentation model. For the diabetic retinopathy dataset, the proposed DCNN classification and grading model achieves 98.63%, 98.61%, 99.58% and 99.09% accuracy, precision, recall and F1-score, respectively. Similarly, for glaucoma, the proposed DCNN classification and grading model achieves 97.81%, 96.67%, 98.32% and 97.49% accuracy, precision, recall and F1-score, respectively. By comparing the performances, the proposed model outperforms existing research.
In real-world graph data, in addition to the class imbalance problem, there also exists topological imbalance. However, existing methods mainly focus on addressing class imbalance or only utilize shallow topological information. To address this issue, we first analyze the impact of graph robustness on label propagation and find that local topological robustness, measured by graph efficiency, plays a crucial role in label propagation. Based on this finding, we propose the GLoFT framework, which redefines topological weights by integrating global influence and local topological robustness, effectively addressing the topological imbalance problem. However, the implementation of GLoFT's core idea relies on high-quality labelled data generated by clustering. In partially labelled scenarios, a decline in pseudo-label quality can significantly weaken its effectiveness and even intensify the class imbalance problem. To cope with this challenge, we propose the fusion-based progressive matching for self-training module (FPMist), which is specifically designed for GLoFT, aiming to enhance the quality of pseudo-labels generated by clustering and improve GLoFT's robustness in partially labelled data settings. Experimental results demonstrate that GLoFT + FPMist outperforms existing methods constructed on multiple GNN backbones and in scenarios with severe class imbalance, showing its superior generalization ability and stability.
In education systems, learners face the challenge of selecting appropriate courses that align with their academic goals or training needs. This paper introduces a hybrid course recommendation system that combines collaborative filtering and knowledge graphs to provide personalised and explainable course suggestions that guide students through their enrollment process. Unlike traditional recommendation methods, which often lack transparency and may offer unsuitable courses for students with different enrollment needs, our system helps prevent redundant suggestions and ensures that students receive appropriate course options based on their preferences and interactions. The system uses a Bernoulli matrix factorisation model to predict the course status, and a filtering method through information extracted from a knowledge graph to provide students with an effective and explainable recommendation. Empirical evaluation of a real-world dataset of 36.6K student records indicates that the proposed system achieves slightly higher levels of precision, recall and F1-score compared to Dirichlet matrix factorisation, naive Bayes collaborative filtering and neural collaborative filtering models. Furthermore, the Bernoulli matrix factorisation model also achieves good performance using the normalised discounted cumulative gain quality measure. The integration of the Bernoulli matrix factorisation model with knowledge graphs offers course recommendations that are accurate, diverse and transparent, improving the learning experience.
Protection of medical images during transmission and storage is critical for ensuring patient confidentiality. To provide secure encryption for healthcare applications, this study proposes the SHA-512 Seeded Multi-Layer Image Encryption (SMLIE) algorithm. It generates a session-specific key by hashing a password and the timestamp of the system using SHA-512, from which the initial states and parameters of the chaotic maps are deterministically derived for reproducibility. The encryption procedure consists of seven layers: Arnold scrambling, logistic map XOR, SHA256 masking, adaptive salt injection, row-column permutation, DWT embedding, and Henon map-based XOR, all of which improve diffusion, entropy, and adaptability to various attacks. A MATLAB graphical user interface (GUI) is used to visualize the encrypted pixel data during validation. The experimental findings indicate a high level of entropy, NPCR, and UACI, demonstrating the effectiveness of the algorithm for encrypting medical images in real time, especially in resource-constrained environments. SMLIE is suitable for real-time healthcare systems in which both privacy and low latency are essential.
Automatic recognition of symbols in computer-aided design (CAD) drawings is a key bottleneck in achieving digital transformation of engineering design. The main challenge is the multi-scale variability of engineering symbols and their complex contextual dependencies. In this study, CADRNet, a framework for engineering drawing symbol recognition based on a CNN-Transformer hybrid architecture, is proposed. The model solves the problems through two core mechanisms: (1) the dynamic feature pyramid network (DFPN), which uses attention mechanism to guide the extraction and integration of multi-scale features to cope with symbol scale variation and (2) the bidirectional cross-architecture adapter (CTAdapter), which realizes bidirectional interaction and fusion between the local texture features extracted by CNN and the global context information captured by transformer. Experiments on the standard CAD symbol recognition benchmark show that CADRNet achieves 0.594, 0.876 and 0.678 in the three key indicators PQ, SQ and RQ of the test set, respectively, surpassing the most advanced CADTransformer model by an absolute improvement of 0.084, 0.080 and 0.037 (that is, a relative improvement of 16.5%, 10.1% and 5.8%, respectively). These results verify the effectiveness and advancement of CADRNet in the field of automated analysis of engineering drawings.
Distributed Graph Neural Networks (GNNs) require efficient handling of both fine-grained memory accesses and cross memory-device communication, particularly when scaling to large graphs. However, existing acceleration solutions fail to adequately address low bandwidth utilization and scalability across different models and graph sizes. In this paper, we present OptGNN, a scalable heterogeneous distributed architecture tailored for GNNs. OptGNN addresses the challenges of fine-grained memory access by incorporating a near-memory processing mechanism, which improves internal bandwidth utilization. To optimize external communication, we introduce data packing and scheduling strategies that enhance cross memory-device data transfer efficiency. OptGNN achieves 5.7x performance improvement over baseline distributed GNN acceleration methods and 1.29x performance improvement over SOTA distributed GNN acceleration architecture CLAY. Additionally, the system is designed to support various GNN models and large-scale graphs while ensuring load balancing and high hardware utilization.
