The Mixture-of-Experts (MoE) architecture has emerged as a promising paradigm for scaling large language models (LLMs) by activating only a sparse subset of experts per input. However, its massive parameter size remains a major obstacle to efficient deployment. Existing pruning methods often ignore two key aspects: the intricate structural dependencies among experts and the heterogeneous importance of different layers. To tackle these issues, we propose C-GNN-PRUNE, a unified and structure-aware compression framework tailored for MoE models. Our method introduces an EntropyGuided Allocation Module that dynamically assigns pruning budgets by leveraging expert activation entropy, enabling adaptive handling of inter-layer heterogeneity. To preserve structural collaboration patterns, we construct an expert interaction graph that fuses functional similarity and routing behavior, and employ a GNN-Based Embedding Module to learn structure-aware expert representations. These embeddings, along with co-activation patterns, are fed into a Community Detection Module to identify expert clusters for structured pruning. Finally, an Activation-Aware Selection Module retains the most critical experts in each community, balancing sparsity and expressiveness. Experiments on multiple open-source MoE models demonstrate that C-GNN-PRUNE consistently outperforms prior methods under various pruning ratios, achieving better trade-offs between compression and accuracy. This framework provides a modular and effective solution for structure-preserving compression of large-scale MoE models.

Evaluating autonomous vehicles’ performance in complex, long-tail traffic scenarios, especially under extreme conditions, often highlights the limitations of existing methods in generating realistic and challenging scenarios, which can affect vehicle safety and reliability. To address these gaps, this paper proposes a Scenario-Adaptive Gradient Adjustment (SAGA) network, an adversarial model specifically designed to generate intricate traffic scenarios that closely mimic real-world dynamics. The SAGA network includes a generator and a victim model, where the generator uses adversarial sequences based on the kinematic bicycle model to simulate dynamic vehicle characteristics and calculate precise gradients. These gradients are then used to perturb the victim model, creating safety-critical scenarios essential for evaluating autonomous vehicle performance, such as emergency evasions and complex intersection navigation. Additionally, we use a k-means++ clustering method tailored to categorize ten types of safety-critical scenarios for autonomous driving. The generated scenarios are comprehensive, diverse, and challenging, providing a robust testing environment for autonomous vehicles. Simulation results demonstrate the SAGA network’s effectiveness, significantly outperforming traditional non-transparent optimization and the KING method. SAGA achieved a 25% higher success rate than non-transparent optimization and a 2% improvement over the KING method in generating complex scenarios, along with an 8% increase in interpretability, reaching 80%. These findings highlight the capability of SAGA-generated scenarios to thoroughly assess autonomous vehicle performance under diverse traffic conditions, ensuring safer operations.

To meet strict Service-Level Objectives (SLO), contemporary Large Language Models (LLMs) decouple the prefill and decoding stages and place them on separate GPUs to mitigate the distinct bottlenecks inherent to each stage. However, the heterogeneity of LLM workloads causes producer-consumer imbalance between the two instance types in such disaggregated architecture. To address this problem, we propose DOPD (Dynamic Optimal Prefill/Decoding), a dynamic LLM inference system that adjusts instance allocations to achieve an optimal prefill-to-decoding (P/D) ratio based on real-time load monitoring. Combined with an appropriate request-scheduling policy, DOPD effectively resolves imbalances between prefill and decoding instances and mitigates resource allocation mismatches due to mixed-length requests under high concurrency. Experimental evaluations show that, compared with vLLM and DistServe (representative aggregation-based and disaggregation-based approaches), DOPD improves overall system goodput by up to$1.5\times$, decreases P90 time-to-first-token (TTFT) by up to 67.5%, and decreases P90 time-per-output-token (TPOT) by up to 22.8%. Furthermore, our dynamic P/D adjustment technique performs proactive reconfiguration based on historical load, achieving over 99% SLO attainment while using fewer additional resources.

The data that records road traffic accidents often hides important information. If we can analyze the main factors that affect the severity of the accident in the data, it has very good practical significance for avoiding the occurrence of major traffic accidents. In this paper, a method called CFS-BAS-BP is proposed to realize the identification of potential traffic accident risk factors based on combinatorial feature selection and bionic neural network. This method combines the advantages of Information Gain, Chi-square test and Random Forest in feature selection to screen the important features affecting traffic accidents to the greatest extent, and filter out irrelevant and redundant features in the data. As a result, 14 features that influence the occurrence and severity of traffic accidents are selected from 25 features in the dataset At the same time, the back propagation neural network optimized by the beetle antennae search algorithm is used to analyze the quantitative relationship between traffic data features and traffic accident risk level, so as to identify the main risk factors affecting the severity of traffic accidents in different scenarios. Through a large number of experiments on the real traffic accident data set STATS19, we prove that CFS-BAS-BP model has a significant effect on the selection of accident-related features, and its accuracy rate of identifying the risk factors affecting the severity of traffic accidents reaches 86.1%.

The availability of high-resolution remote sensing images provides rich geographical and environmental insights but also presents challenges due to complex spatial distributions and diverse semantics. To address the challenge of insufficient fine-grained semantic understanding in change description models for remote sensing images, we introduce a novel framework combining Global Semantic Extraction (GSE) and Adaptive Cross-Semantic Learning (ACSL) to enhance change captioning. The framework includes components: (1) GSE for global scene understanding; (2) ACSL to refine feature interactions and distinguish meaningful changes from irrelevant variations; and (3) a multi-layer Transformer decoder that dynamically generates accurate, context-aware change descriptions. Evaluation on the Dubai-CC and LEVIR-CC datasets shows the framework outperforms current methods, offering a more adaptive and precise solution for remote sensing change captioning.

In edge computing environments, service reliability is often threatened by the sudden failure of edge nodes due to harsh deployment conditions, leading to task interruption and performance degradation. To address this challenge, EFTR-HGNet is proposed as a novel task rescheduling framework tailored for edge-fault scenarios. It leverages heterogeneous graph neural networks with a Transformer-based architecture to achieve cost-efficient task migration and adaptive decision making. Specifically, the rescheduling problem is formulated as a Markov Decision Process (MDP), and a three-dimensional fault-aware state representation that jointly encodes task attributes, resource availability, and dynamic failure status is introduced. To model the complex relationships between failed tasks and heterogeneous edge resources, a heterogeneous Transformer (HG-Trans) network is designed, which performs two-stage embedding over the constructed graph, enabling context-aware rescheduling decisions to be made by the agent. By optimizing both the policy and value functions within an Actor-Critic reinforcement learning framework, our method achieves a favorable balance between minimizing the overall Makespan and maximizing the task rescheduling success rate. Evaluated against strong baselines like HEFT, TDCA, and FixDoc, EFTR-HGNet demonstrated superior performance, achieving a Makespan reduction of at least 11.11% and a 4.20% increase in task rescheduling success. These results highlight its robustness and practical potential for fault-prone edge computing systems.

The surge in the number of e-bikes also brings certain problems of traffic road order and safety management. Analyzing and mining the travel routes of e-bikes to discover the valuable patterns latent in large-scale trajectories will help the traffic department to manage e-bikes. For this reason, this paper proposes a frequent pattern mining method for trajectories based on deep clustering, which shows the travel patterns of e-bikes through frequent trajectory patterns in order to improve the transportation service capacity. In the feature extraction stage, the method transforms the trajectory point sequence into a raster sequence, and obtains the potential vectors of e-bike trajectories based on the autoencoder; in the clustering stage, it integrates the deep learning model and the improved clustering algorithm by adding the chameleon model to mine the frequent pattern of trajectory vectors, and completes the unsupervised clustering of deep learning by updating the parameters and clustering centers in the feature extraction stage. Finally, the effectiveness and application value of our method are proved by a large number of experiments.

With the rapid development of big data and artificial intelligence technologies, a variety of high-performance plat-forms and applications in data centers have significantly increased their energy consumption and operating costs. Energy consumption prediction plays an important role in the optimization and management of energy consumption in data centers. Due to the complexity of power system, there are many complex correlations between various energy consumption parameters and prediction results. In order to improve the accuracy of energy consumption prediction, in this paper, a new prediction model based on BiLSTM is proposed, name SLCOA-BiLSTM (Saltation learning-based coati optimization algorithm and Bi-directionallong short-term memory). Aiming at the minimum prediction error, the model adopts the improved coati optimization algorithm (SLCOA) to optimize the BiLSTM network's learning rate, lstm_layers number, number of units and dropout. In the process of optimization, in order to solve the shortcomings of coati optimization algorithm, Sobol sequence initialization and Salation learning are improved. The experiments are carried out based on the real-time energy consumption data of a data center, and the results show that the model is superior to the relative comparison model in terms of prediction error. Since SLCOA-BiLSTM predicts are stable and accurate, it can be used as an effective tool to solve energy consumption problems.

There exist some challenges in accurately capturing temporal change information and efficiently aggregating multi-level information in the field of remote sensing change detection. In order to expand the detection’s receptive field and fully fuse complementary information across different hierarchical levels, we propose a network based on progressive feature interaction and cross-scale feature fusion(PFCF-Net). Specifically, PFCF-Net adopts a naive backbone network ResNet-18 for efficient feature extraction. The progressive feature interaction module utilizes dilated convolutions with different dilation rates to capture feature changes at various scales, effectively capturing a wide range of changes from macroscopic to detailed levels in remote sensing images. The cross-scale feature fusion module improves cross-attention mechanism with the assistance of disparity guidance and peer guidance. It reduces semantic ambiguity and spatial detail loss in detected change objects, allowing the model to more effectively focus its detection efforts on regions of interest. Through extensive comparative experiments on three benchmark datasets, PFCF-Net has surpassed several state-of-the-art change detection methods in terms of accuracy.

With the development of Internet of Vehicles(IoV) technology and the popularization of intelligent vehicles, a large amount of vehicle data is generated and accumulated. Due to the privacy and sensitivity of vehicle datas, the emergence of Federated Learning(FL) technology can effectively protect user privacy. FL allows model training without the need to share raw data, protecting vehicle user data by sharing model parameters. However, within conventional FL methodologies, vehicle users are required to undertake significant computational tasks, and communication with cloud servers demands considerable time. Inspired by edge computing, we introduce a cloud-edge-end two-tier architecture based on federated learning(CEE-FLA).This structure transfers the model aggregation process originally completed by the cloud server to the edge server to reduce the number of communications with the cloud server, thereby improving the accuracy and efficiency of data sharing. Considering that model training efficiency is affected by vehicle performance, we improve the traditional FL algorithm in CEE-FLA and adopt a K-Vehicle random scheduling strategy to minimize the impact of vehicle performance differences on model convergence and data sharing efficiency. The experimental results demonstrate that our proposed CEE-FAL offers improvements in both accuracy and efficiency compared to the traditional cloud-end single-tier. Additionally, as the number of iterations between vehicles and edge servers increases, the model’s convergence speed also improves.

Mobile edge computing can effectively meets the resource requirements of complex vehicle applications with high computing power and low delay. When a vehicle terminal task is offloaded to the edge servers, most works focus on the demands of the vehicle terminal and the service quality of the offloaded task when selecting an optimal resource allocation schemes. But it also creates a new problem of the abuse of edge service resources. Aiming at the low utilization of edge resources in the Internet of vehicles, a resource allocation method based on the bundled two-way auction model is proposed in this paper. From the perspective of game, the edge resources are firstly bundled, and then the resource pricing, quotation and inquiry strategy and payment rules in the auction are analyzed. At the same time, the value gain of the vehicle application and the cost consumption of the edge resources in the auction are used as the targets to optimize the resource allocation scheme of edge service system. The experimental results show that the method proposed in this paper has higher resource utilization and faster convergence than the existing methods.

Mobile edge computing can expand the limited computing power of on-board equipment, which is an effective means to assist vehicles to realize complex applications. However, the task offloading process brings a certain delay. How to achieve faster offloading to reduce the delay of vehicle tasks is crucial for delay-sensitive vehicle tasks. The deployment of edge servers is the basis of vehicle computing task offloading, and an efficient deployment method of edge servers can effectively meet the low latency of mobile vehicle access. Therefore, in order to minimize the average access delay between AP points and edge servers, this paper establishes a deployment model of edge servers, and proposes a mobile edge server deployment method based on the K-means algorithm and the genetic algorithm, called KGA, which fully considers the position relationship among vehicles, APs and edge servers and the limitation of wireless transmission bandwidth to determine the deployment location of edge servers and optimize communication delay of all vehicle tasks. Finally, a large number of simulation experiments under different scenarios were completed. The experimental results show that the proposed KGA method can effectively reduce the delay under the premise of the edge server capacity limitation, and its effect is better than several existing representative algorithms.

High radix SRT division plays an important role in contemporary microprocessors as the quotient digit selection tables effectively reduce the computation complexity of the quotient digits. The quotient digit selection table is constructed according to the rounded lower and upper bounds of the overlapping regions in the traditional method. The table construction process lacks in mathematical rigor and consequently is susceptible to error. This paper proposes an algebraic method for computing the quotient digit selection tables. We characterize the quotient digit selection functions to construct quotient digit selection tables required for SRT division and SRT square root with any valid redundancy. The functions include the maximum and minimum legal quotient digit selections. We compute the truncations of the remainder$p$and divisor$d$when$d \in [1,2)$. We implement procedures to compute quotient digit selection tables by our functions. The computation of the quotient digit selection table for the case radix-4 with the quotient digit set$[-2,2]$is presented by using the minimum quotient digit selection function. Our functions can compute quotient digit selection tables in the design phase of SRT division and square root by given radix and redundancy.

Automatic extraction of text information plays an important role in machine translation, knowledge mapping and other fields. In recent years, with the rapid development of computer technology and the popularization of Internet application, the resources acquired by people through the Internet show explosive growth. Facing the massive information resources, how to extract the required information quickly and effectively and convert it into structured data has become a hot topic of current research. Based on this, this paper proposes a automatic topic-oriented text extraction method combining BiLSTM and CRF models. This method firstly establishes the text extraction topic, then carries on the automatic entity recognition to the data related topic, finally forms the standard structured data, thereby realizes the unstructured text data to the specific structure of the information block, which lays the foundation for knowledge mining. Taking the data collected in ACL 2018 Chinese NER as the test data, the precision of our algorithm is 95.74%. Compared with the traditional neural network method, our text information extraction method can effectively identify more entity information in the text data, and improve its effect in practical application.

Cloud services request lower cost compared to traditional software of self-purchased infrastructure due to the characteristics of on-demand resource provisioning and pay-as-you-go mode. Current enterprises compact their business software as services into cloud platform to users. In the cloud services market, service providers attempt to make more profits from their services, while users hope to choose low-cost services with high-quality. The conflict of interests between users and service providers is an important challenge for the booming cloud service market. This article characterizes this application problem formally based on a utility game model of service providers and users. In the model, QoS is considered as the basis for determining the utilities of both parties from an economic point of view. By analyzing the behaviors of users and service providers, we introduce the concept of reputation cost for the first time in the model and find a QoS solution that balances the utilities of users and service providers in service transactions. In such a balance, any change in either party's strategy will result in a loss of utility. And then a QoS optimization method is designed to obtain a near-optimal QoS solution for a tradeoff between user satisfaction and provider profit. Extensive simulation experiments are conducted to substantiate the effectiveness of our method. The results are applicable to win-win service applications between service providers and users.

In the age of the Internet of everything, the Edge-Cloud collaborative service support has become a very promising development direction in the application field of the Internet of things. However, when various service components are deployed both on the cloud and the edge, the subsidence and decentralization of computing resources also have a great impact on the performance of composite services. This paper proposes an optimal composite service selection model based on Petri nets. In order to compare the composite service paths, this paper adopts a dynamic evaluation model of composite service quality based on Petri nets. Firstly, all kinds of Petri net models for the implementation structure of composite services are constructed. And then the QoS computing rules corresponding to these structures are given based on the QoS of each service component on the edge of the cloud. Finally, the dynamic execution process of each composite service with a feasible path is simulated through Petri net, and the overall performance of the service is evaluated based on the dynamic simulation of its Petri net model, finally, the optimal service path is selected among the combined services that meet the user's requirements.. Through case analysis and comparison, the feasibility and effectiveness of the method are verified by an example analysis.

With the rapid development of Internet, various forms of online learning platforms have emerged. As a major form of knowledge presentation, a number of electronic courseware (e-courseware) have been uploaded to these platforms for users to learn and share. Extracting main contents of an e-courseware document to form a knowledge framework is of great helpful for learners to select and utilize some courseware documents for their self-learning process. But at present, there are few studies concentrated on the information extraction for e-courseware. In this paper, a knowledge mining algorithm for e-courseware is proposed to facilitate learners quickly grasp main knowledge points of the courseware. In order to effectively organize the knowledge framework, this algorithm firstly uses the IRAKE algorithm to get some key phrases in each page of the e-courseware, and then uses the MMS process to filter the phrases with similar meanings, finally uses the query likelihood model to retrieve the sentences related to these key phrases in all the pages of the e-courseware document. The algorithm has good performance in mining the knowledge covered by the e-courseware.

Abstract The research on personalized recommendation of Web services plays an important role in the field of Web services technology applications. Fortunately, not all users have completely different service preferences. Due to the same application scenarios and personal interests, some users have the same preferences for certain types of Web services. This paper explores the problem of user clustering in the service environment, grouping users according to their service preferences. It helps service providers to identify and characterize the preferences of similar users and provide them with customized services. We propose two combination-based clustering algorithms which make full use of the advantages of the K-means algorithm and the affinity propagation algorithm. In addition, a three-stage clustering process is elaborated to improve the accuracy of user clustering. To reduce the time complexity of the algorithms, we create a parallel execution model of the algorithms implemented by a higher-order MapReduce sequence linking technology. Extensive experiments on simulated datasets and real datasets are performed on the comparisons between the proposed algorithms and the other combination-based clustering algorithms. The experimental results substantiate that the proposed algorithms can effectively distinguish user group with different preferences.

As the technology of cloud computing becomes mature, the service idea of IT resources has been promoted rapidly. The various available cloud services have different quality of service (QoS). Different users have different QoS requirements for services. Finding the best QoS strategy which a user is satisfied with is a multi-criteria NP-hard problem. The completed theory and applied research have proved that the swarm intelligence method can effectively solve most multiobjective optimization problems. Therefore, a Pareto optimal solution for QoS strategy can be found by an evolutionary algorithm based on swarm intelligence. However, so far, only few solutions based on these approaches have been proposed and there exists no comparative study published to date. This motivated us to perform an analysis of state of the art evolutionary algorithm based on swarm intelligence. By a large number of contrast experiments under different circumstances, three QoS evolutionary algorithms based on swarm intelligence are analyzed and compared.

Cloud services typically compose of multiple distributed software components that communicate with each other through web service interfaces in the cloud environments. During their long time running, the accumulation of cloud software internal errors or large consumption of computing resources will very likely lead to software aging problems. In order to solve this problem, software rejuvenation technology is proposed to prevent them from causing more serious failures by restarting the services running. In the research field of software aging and rejuvenation for cloud services, how to accurately predict the cloud resource consumption in the aging software system for determining suitable time to perform rejuvenation is a significant and indispensable issue. In this paper, a novel hybrid aging prediction model named CSSAP is proposed, which well integrates the Autoregressive Integrated Moving Average (ARIMA) model and Long Short Term Memory (LSTM) model for better fitting the linear pattern and mining the nonlinear relationship in the time series of computing resource usage data for cloud services. The experiments results show that through such hybrid and unified time series analysis, our CASSP prediction method has 4% to 71% improvements in MAE evaluation criteria and 6% to 66% improvements in RMSE evaluation criteria under different time series scenarios compared with single model used, that is, the more accurate and more comprehensive aging prediction results achieved by CSSAP is definitely conducive to perform more effective and more efficient software aging and rejuvenation for cloud services.

The software sharing platform in the Internet provides great convenience for the promotion, application and communication of software (especially source software). But there inevitably exists the problem of software quality on the open Internet platform. How the users choose software to download and use becomes a new challenge for software sharing platforms. Aimed at the above problems and challenges, the internal relations between the data collected on platform and experience of user are analyzed. And then a software popularity recommendation method based on evaluation model is presented. The method constructs two evaluation indexes based on the collected data on the platform, including attention-degree and satisfaction-degree; solves the problem of small sample data's influence on the accuracy of evaluation model by using the Wilson interval model and makes a tradeoff between the recommendation results of old and new software by using Newton cooling law. The experimental results show that the software popularity recommendation method based on evaluation model helps users to screen for software, which can effectively improve the service performance of software sharing platform.

In order to meet the demand of people's travel, all kinds of hailing-taxi apps have emerged. Although the advent of these apps brings people convenience greatly, their shortcomings and deficiencies are also increasingly apparent. Due to the subjectivity of the taxi drivers, some drivers may pick up a passenger without considering the problem of traffic congestion, which decrease the efficiency of taxis. So an efficient coordination of taxi network at a large scale becomes a new challenge in the intelligent transportation. In order to better solve the insufficiency of taxi-hailing apps, strengthen the passengers' experience and improve the efficiency of taxi service, an intelligent recommendation service for taxi based on real-time traffic in the road network is designed. And then a global optimization algorithm based on simulated annealing particle swarm optimization is presented to search a vacant taxi which can reach the location of the passenger in the shortest time. The comparative experiments with other algorithms prove that the algorithm has certain effectiveness and efficiency.

With the development of web service technology, identifying and discovering the users with similar preferences have an important significance to service selection and service optimization in the service environment. In order to divide the users into groups based on their preference similarity in the process of service selection, a combination-based clustering algorithm, named AAK, is presented in this paper. The method combines the K-means algorithm with the Affinity Propagation (AP) algorithm to cluster the users with similar preferences. In the clustering process, the algorithm makes full use of the advantages of the two algorithms, including the high partition accuracy of K-means algorithm and the independence in the prior knowledge of AP algorithm, which breaks the limitation of using a single clustering algorithm. Then a parallel execution model of the algorithm is built and implemented by a high order MapReduce sequence linking technology. Finally AAK algorithm is compared with its serial model and the other combination-based clustering methods on Matlab platform and Hadoop platform. The experimental results show that AAK algorithm can be applied to distinguish user group with different preferences and has a good effectiveness and efficiency.