In response to challenges in natural language instruction parsing, dynamic task scaling, and rigid optimization strategies in urban multi-UAV emergency delivery, this paper proposes a large language model (LLM)-driven multi-agent UAV deployment framework. The LLM serves as a cognitive scheduler for task understanding, structuring, and strategy selection, while Google OR-Tools performs vehicle routing optimization under capacity and priority constraints to ensure feasibility and efficiency. The system adopts a closed-loop multi-agent architecture including intent parsing, task decomposition, deployment optimization, and execution monitoring. Experiments under different task scales compare heuristic, pure LLM, and pure OR-Tools baselines. Results show that the proposed method maintains 100% task completion while reducing total path cost and improving high-priority task satisfaction, demonstrating the scalability and engineering feasibility of the hybrid LLMoptimization architecture.

Unmanned aerial vehicles (UAVs) are promising mobile platforms for enhancing sensing performance in integrated sensing and communication (ISAC) systems. However, many existing UAV-enabled ISAC designs remain communication-centric or rely on transmit-side surrogate sensing metrics. This paper proposes a Cramér–Rao bound (CRB)-driven, sensing-centric multi-antenna UAV-ISAC framework that exploits UAV mobility and multi-antenna transmission to enhance angle-estimation performance. The transmit beamforming and UAV trajectory are jointly optimized to maximize the time-averaged sum of reciprocal angular CRBs over multiple targets. The formulation accounts for bounded CSI uncertainty and receiver-side sensing-interference cancellation, while enforcing worst-case communication SINR, per-target per-slot sensing-accuracy, transmit-power, and UAV-mobility constraints. To address the resulting nonconvex problem with coupled transmit and trajectory variables, we develop an SDP–SOCP alternating optimization (SSAO) framework. For transmit optimization, tractable LMI reformulations are derived for the worst-case SINR and CRB-related constraints, with a sufficient robust-feasibility certificate for the rank-one realization. For trajectory optimization, candidate trajectories are generated via a trust-region-controlled SOC approximation and accepted only after original-constraint and objective-improvement checks. Numerical results demonstrate that the proposed CRB-driven design consistently outperforms representative baselines and reduces the average root CRB by up to 28.1% relative to the beampattern-gain-based benchmark under the considered setup.

Multimodal medical data provides a wide and real basis for disease diagnosis. Computer-aided diagnosis (CAD) powered by artificial intelligence (AI) is becoming increasingly prominent in disease diagnosis. CAD for multimodal medical data requires addressing the issues of data fusion and prediction. Traditionally, the prediction performance of CAD models has not been good enough due to the complicated dimensionality reduction. Therefore, this paper proposes a fusion and prediction model—EPGC—for multimodal medical data based on graph neural networks. Firstly, we select features from unstructured multimodal medical data and quantify them. Then, we transform the multimodal medical data into a graph data structure by establishing each patient as a node, and establishing edges based on the similarity of features between the patients. Normalization of data is also essential in this process. Finally, we build a node prediction model based on graph neural networks and predict the node classification, which predicts the patients’ diseases. The model is validated on two publicly available datasets of heart diseases. Compared to the existing models that typically involve dimensionality reduction, classification, or the establishment of complex deep learning networks, the proposed model achieves outstanding results with the experimental dataset. This demonstrates that the fusion and diagnosis of multimodal data can be effectively achieved without dimension reduction or intricate deep learning networks. We take pride in exploring unstructured multimodal medical data using deep learning and hope to make breakthroughs in various fields.

Unmanned aerial vehicles (UAVs) have a great potential for assigning search and rescue operations in emergency scenarios. However, emergency scenarios are complex and unknown, regarding UAVs to reschedule to effectively adapt to the changing environment, and existing literature addressing this challenge is limited. To address this open problem, we consider a task rescheduling problem with uncertainties such as task insertion, edge computing node (ECN) destruction, and parameter fluctuation in UAV-assisted emergency networks. The goal is to minimize the fine-grained makespan, defined as the ratio of makespan to ECNs idle time, that simultaneously characterizes the optimization of rescheduling efficiency and ECNs utilization. To address the problem, we propose an asynchronous shuffled frog-leaping with feasible Jaya (ASFJ) algorithm. In ASFJ, an asynchronous shuffled frog-leaping method independently evolves memeplexes, thereby avoiding forced information coverage. Two feasible local search operators promote the search capability and feasibility of the algorithm. Finally, we verify the advantages of the ASFJ in terms of makespan, effectiveness, and fine-grained makespan. ASFJ can save 3.83ms makespan and outperform 11.2% fine-grain makespan in insertion rescheduling. The effectiveness of destruction rescheduling is improved by at least 16%.

The problem that the texture details of low-resolution (LR) digital images of ancient murals are ambiguous persists. To solve this problem, this study proposes a super-resolution (SR) reconstruction method for fuzzy murals based on a generative adversarial network with self-attention (SA). The network uses a blur kernel and realistic noise data to add blur and noise, respectively, to a high-resolution (HR) image to obtain an original LR image. Then, a feature image with the same size as that of the input image is obtained through a SA module. Finally, the feature information extracted from the image is input into the high-resolution image space by using a subpixel convolution layer to realize the image enlargement process from an LR to an HR. Experiments evaluate the proposed approach both objectively and subjectively. The objective evaluation results show that compared with other SR reconstruction algorithms, the proposed algorithm's peak signal-to-noise ratio (PSNR) is increased by 0.04 to 3.78 dB on average, and its structural similarity is increased by 0.002 to 0.191. A subjective perception evaluation shows that the developed algorithm can better reconstruct the texture details of murals, thus better meeting the visual perception needs of the public. The method proposed in this study can satisfactorily reconstruct the texture details of murals, which may provide technical guidance for the development of mural protection plans. Furthermore, it may be of certain practical significance for the SR reconstruction of ancient murals.

In dynamic and unknown emergency networks, (unmanned aerial vehicle) UAV-assisted task scheduling is an important and efficient technique. However, the inevitable uncertainty makes the predetermined schedule decision unfeasible. In this paper, we investigate the effects of parameter fluctuation and unpredictable edge computing node (ECN) failure. We formulate the task rescheduling problem to effectively deal with uncertainty, and design an asynchronous shuffled frog-leaping with jaya (ASJ) algorithm to minimize the makespan, in which the jaya algorithm asynchronously evolves the memeplexes of the shuffled frog-leaping method. To verify the effectiveness of the proposed ASJ, experiments are conducted to compare it with two comparison algorithms under deterministic and uncertain scenarios. The results demonstrate the superiority of the ASJ in terms of makespan.

Multi-user time series data is critically important in predicting future user behavior across various fields, including communication and finance. However, due to data skewness, user heterogeneity, and time shift, it becomes challenging to predict multiple types of user time series effectively. Currently, there are few efficient time series prediction methods that address these issues. In this paper, we propose a dynamic time warping based radial basis function neural network model for multi-user time series prediction, named DTW-RBFNN. To address data skewness, it introduces an effective preprocessing function of logarithmic power function to adjust the users' range and distribution. K-means based on DTW and DBA was utilized to eliminate the time-shift issue and obtain proper centers for the RBFNN model. Finally, the DTW-RBFNN model overcomes user heterogeneity and directly predicts the time series of different users in a unified manner. Our experimental evaluation on real-world user communication traffic data and bank user deposit data demonstrates the superiority of the proposed DTW-RBFNN model over the traditional fixed effects model and deep neural network model LSTM.

The proliferation of the Internet of Things (IoT) and widespread use of devices with sensing, computing, and communication capabilities have motivated intelligent applications empowered by artificial intelligence. Classical artificial intelligence algorithms require centralized data collection and processing, which are challenging in realistic intelligent IoT applications due to growing data privacy concerns and distributed datasets. Federated Learning (FL) has emerged as a privacy-preserving distributed learning framework, which enables IoT devices to train global models through sharing model parameters. However, inefficiency due to frequent parameter transmissions significantly reduces FL performance. Existing acceleration algorithms consist of two main types including local update and parameter compression, which considers the trade-offs between communication and computation/precision, respectively. Jointly considering these two trade-offs and adaptively balancing their impacts on convergence have remained unresolved. To solve the problem, this paper proposes a novel efficient adaptive federated optimization (FedEAFO) algorithm to improve the efficiency of FL, which minimizes the learning error via jointly considering two variables including local update and parameter compression. The FedEAFO enables FL to adaptively adjust two variables and balance trade-offs among computation, communication, and precision. The experiment results illustrate that compared with state-of-the-art algorithms, the FedEAFO can achieve higher accuracies faster.

Federated Learning (FL) has emerged as a privacy-preserving distributed learning framework which enables IoT devices to collaboratively train machine learning models vi-a sharing model parameters. However, inefficiency due to frequent parameters transmissions significantly reduces FL performance. Existing acceleration algorithms for speeding up FL training consist of two main types including local update and parameter compression which consider the trade-offs between communication and computation/precision respectively. Jointly considering these two trade-offs and adaptively balancing their impacts on convergence have remained unresolved. To solve the problem, we propose an efficient adaptive federated optimization (EAFO) algorithm to improve the efficiency of FL in resource-constrained IoT environments, which minimizes the learning error by the joint consideration of two variables consisting of the local update and parameter compression. The EAFO enables FL to adaptively adjust two variables and balance trade-offs among computation, communication, and precision. The experiment results illustrate the high effective-ness of the proposed EAFO algorithm, which can achieve higher accuracies faster compared with the state-of-the-art algorithms.

Introduction: A twin pregnancy involving a hydatidiform mole (HM) coexisting with a developing fetus is an extremely rare obstetric complication, which typically presents as a complete hydatidiform mole with a coexisting fetus (CHMCF) or a partial hydatidiform mole with a coexisting fetus (PHMCF). Case presentation: A 26-year-old woman was admitted to our hospital due to a small volume of vaginal bleeding during the 31st week of pregnancy. The patient was previously healthy, and an intrauterine singleton pregnancy was detected by ultrasound on day 46 of gestation; however, bunch-of-grapes sign was observed in the uterine cavity at 24 weeks. The patient was subsequently diagnosed with CHMCF. As the patient insisted on continuing her pregnancy, she underwent hospital monitoring. Vaginal bleeding occurred in the 33rd week again and received a course of betamethasone, then continued pregnancy after bleeding stopped spontaneously. In the 37th week, a male infant weighing 3090 g was delivered by cesarean section, with an Apgar score of 10 at 1 min and a karyotype of 46XY. Placental pathology confirmed the diagnosis of a complete hydatid tumor. Conclusion: In this report, a case of CHMCF was maintained by monitoring of blood pressure, thyroid function, human chorionic gonadotrophin, and fetal condition during pregnancy. A live newborn was delivered by cesarean section. CHMCF is a clinically rare disease with high risks; thus, it should be diagnosed carefully using several tools, including ultrasound, magnetic resonance imaging, and karyotype analysis and dynamically monitored if the patient decides to continue the pregnancy.

Node-based link prediction methods have occupied a dominant position in the graph link prediction task. These methods commonly aggregate node features from the subgraph to generate the potential link representation. However, in constructing subgraphs, these methods extract each node's local neighborhood from the target node pair separately without considering the correlation between them and the whole node pair. As a result, many nodes in the subgraph may have little contribution to predicting the potential edge. Aggregating these node features will reduce the model's accuracy and efficiency. In addition, these methods indirectly represent the potential link by the node embeddings in the subgraph. We argue that this formalism is not the best choice for link prediction. In this paper, we propose a relation-based link prediction neural network named RelpNet, which aggregates edge features along the structural interactions between two target nodes and directly represents their relationship. RelpNet first extracts paths between the target node pair as structural interactions, which have strong correlations with the whole node pair and fewer nodes and edges than node-based methods' subgraph. To aggregate edge embeddings along the links between edges, we propose transforming the paths into a line graph. Then, the Tree-LSTM model is adopted to transfer and aggregate the node embeddings in the line graph as a comprehensive representation of the target node pair. We evaluate RelpNet on 7 benchmark datasets against 15 popular and state-of-the-art approaches, and the results demonstrate its significant superiority and high training efficiency.

Graph Convolutional Networks (GCNs) have achieved much success in various graph learning tasks. However, as the number of layers increases, the smoothing of GCNs will over-mix the neighbors’ information, leading output towards space with low expressivity. It is known as the over-smoothing issue. Although several works have refined deep GCNs by optimizing network structure, receptive field, and topology, the over-smoothing issue cannot be completely avoided. In this paper, we propose a recurrent neural network framework for learning graph representation while avoiding over-smoothing effectively, which is the tree-structure aggregation and optimization framework named Treeago. Treeago firstly transforms the irregularly distributed graph into sequential trees. Then, Treeago adopts Tree-LSTM with attention to aggregate important neighbors’ feature information to the graph representation. Tree-LSTM with attention can prevent the mixing of noise neighbors’ information to avoid the over-smoothing issue. Finally, Treeago uses an edge pruning optimization framework based on reinforcement learning to enhance the model’s performance further. Experimental results on multiple real-world datasets show that Treeago effectively avoids over-smoothing and yields state-of-the-art results.

Node-based link prediction methods have occupied a dominant position in the graph link prediction task. These methods commonly aggregate node features from the subgraph to generate the potential link representation. However, in constructing subgraphs, these methods extract each node's local neighborhood from the target node pair separately without considering the correlation between them and the whole node pair. As a result, many nodes in the subgraph may have little contribution to predicting the potential edge. Aggregating these node features will reduce the model's accuracy and efficiency. In addition, these methods indirectly represent the potential link by the node embeddings in the subgraph. We argue that this formalism is not the best choice for link prediction. In this paper, we propose a relation-based link prediction neural network named RelpNet, which aggregates edge features along the structural interactions between two target nodes and directly represents their relationship. RelpNet first extracts paths between the target node pair as structural interactions, which have strong correlations with the whole node pair and fewer nodes and edges than node-based methods' subgraph. To aggregate edge embeddings along the links between edges, we propose transforming the paths into a line graph. Then, the Tree-LSTM model is adopted to transfer and aggregate the node embeddings in the line graph as a comprehensive representation of the target node pair. We evaluate RelpNet on 7 benchmark datasets against 15 popular and state-of-the-art approaches, and the results demonstrate its significant superiority and high training efficiency.

Semi-supervised short text classification is a challenging problem due to the sparsity and limited labeled data. Due to the lack of labeled data, many models focus on the generation of text samples, which is cumbersome and has poor scalability. To overcome this deficiency, in this paper, we propose a Self-Training Text method based on Graph Convolutional Networks (ST-Text-GCN). Differently from the previous literature, our self-training method is convenient. The labeled information is propagated to target samples along the structure of the manifold, instead of introducing the extra knowledge. Specifically, instead of adding text training samples, our method adds keywords to training set. The model will calculate the confidence of each word. Confidence indicates the degree of ambiguity of a word. Some words with high confidence are automatically marked as pseudo-labeled data. Meanwhile, word confidence is added to the calculation of the edge weights of the graph to reduce the classification error caused by word ambiguity. Our method makes full use of the keywords in short texts when labeled data is scarce. Extensive experimental results have demonstrated that our proposed method outperforms state-of-the-art models on multiple benchmark datasets.

In recent years, Delay Tolerant Networks (DTN) have received more and more attention. At the same time, several existing DTN routing algorithms generally have disadvantages such as poor scalability and inability to perceive changes in the network environment. This paper proposes an AdaptiveSpray routing algorithm. The algorithm can dynamically control the initial maximum message copy number according to the cache occupancy rate of the node itself, and the cache occupancy rate is added as an impact factor to the calculation of the probability of each node meeting the destination node. In the forwarding phase, the node will first compare the meeting probability of itself and the meeting node to the destination node, and then choose different forwarding strategies. The simulation shows that the AdaptiveSpray algorithm proposed in this paper has obvious advantages compared with the existing routing algorithms in terms of message delivery rate and average delay.

The proliferation of the Internet of Things (IoT) and widespread use of devices with sensing, computing, and communication capabilities have motivated intelligent applications empowered by artificial intelligence. The classical artificial intelligence algorithms require centralized data collection and processing which are challenging in realistic intelligent IoT applications due to growing data privacy concerns and distributed datasets. Federated Learning (FL) has emerged as a distributed privacy-preserving learning framework that enables IoT devices to train global model through sharing model parameters. However, inefficiency due to frequent parameters transmissions significantly reduce FL performance. Existing acceleration algorithms consist of two main type including local update considering trade-offs between communication and computation and parameter compression considering trade-offs between communication and precision. Jointly considering these two trade-offs and adaptively balancing their impacts on convergence have remained unresolved. To solve the problem, this paper proposes a novel efficient adaptive federated optimization (EAFO) algorithm to improve efficiency of FL, which minimizes the learning error via jointly considering two variables including local update and parameter compression and enables FL to adaptively adjust the two variables and balance trade-offs among computation, communication and precision. The experiment results illustrate that comparing with state-of-the-art algorithms, the proposed EAFO can achieve higher accuracies faster.

As a technique to buffer the conflict between computation-intensive tasks and capability-limited devices, unmanned aerial vehicle-enabled mobile edge computing (UAV-MEC) has been witnessed as a promising approach. In this paper, we consider an optimization of dependent and independent subtasks hybrid offloading to maximize the utility, which is decided by the weighted sum of makespan gain and energy consumption gain. Firstly, we design a two-tier UAV-assisted MEC offloading architecture to obtain efficient and cooperative communication between the ground command centers and the UAVs. Secondly, we develop a flexible-hybrid subtask offloading scheme(F-HSO) to obtain the optimal solution of the problem in a parallel and sequential manner. Numerical results show that our algorithm can significantly outperform the other representative benchmarks in utility.

As promising privacy-preserving machine learning technology, federated learning enables multiple clients to train the joint global model via sharing model parameters. However, inefficiency and vulnerability to poisoning attacks significantly reduce federated learning performance. To solve the aforementioned issues, we propose a dynamic asynchronous anti poisoning federated deep learning framework to pursue both efficiency and security. This paper proposes a lightweight dynamic asynchronous algorithm considering the averaging frequency control and parameter selection for federated learning to speed up model averaging and improve efficiency, which enables federated learning to adaptively remove the stragglers with low computing power, bad channel conditions, or anomalous parameters. In addition, a novel local reliability mutual evaluation mechanism is presented to enhance the security of poisoning attacks, which enables federated learning to detect the anomalous parameter of poisoning attacks and adjust the weight proportion of in model aggregation based on evaluation score. The experiment results on three datasets illustrate that our design can reduce the training time by 30% and is robust to the representative poisoning attacks significantly, confirming the applicability of our scheme.

Cerebral stroke (CS) is a heterogeneous syndrome caused by multiple disease mechanisms. Ischemic stroke (IS) is a subtype of CS that causes a disruption of cerebral blood flow with subsequent tissue damage. Noncontrast computer tomography (NCCT) is one of the most important IS detection methods. It is difficult to select the features of IS CT within computational image analysis. In this paper, we propose AC-YOLOv5, which is an improved detection algorithm for IS. The algorithm amplifies the features of IS via an NCCT image based on adaptive local region contrast enhancement, which then detects the region of interest via YOLOv5, which is one of the best detection algorithms at present. The proposed algorithm was tested on two datasets, and seven control group experiments were added, including popular detection algorithms at present and other detection algorithms based on image enhancement. The experimental results show that the proposed algorithm has a high accuracy (94.1% and 91.7%) and recall (85.3% and 88.6%) rate; the recall result is especially notable. This proves the excellent performance of the accuracy, robustness, and generalizability of the algorithm.