学术视点
| 来源:【字号:大 中 小】
题目:GOMOS: A GNN-Assisted Network-Media Integrated Optimization Framework for Massive Mobile Live Streaming(GOMOS:面向大规模移动直播的基于图神经网络的网络与媒体协同优化框架)
作者:F. Feng, Y. Zhang, Z. Wang, T. Bi, M. Feng and T. Jiang
来源:IEEE Transactions on Cognitive Communications and Networking(认知通信与网络汇刊), vol. 11, no. 5, pp. 3333-3345.
摘要:In this paper, a network-media integrated optimization framework assisted by a graph neural network (GNN) named GOMOS is proposed, aiming to address the challenge of reducing overall live latency and avoiding rebuffering events in massive mobile live streaming (MMLS) systems. Compared to existing frameworks that optimize latency from either the media or network perspective, GOMOS profoundly considers the mutual impact between them, and jointly optimizes both wired and wireless networks with three steps. Firstly, a bi-directional link traversal algorithm, adopting a customized TCP traversal packet (TTP), is designed to collect and disseminate the real-time link status of wired networks. Secondly, using graph representation to comprehensively consider the characteristics of media and networks, a gated GNN (GGNN) model is utilized to predict the latency of transmitting MMLS flows with wired links. Thirdly, given the predicted wired transmission latency, the resource allocation of wireless networks is formulated as a cross-layer convex optimization problem, ensuring timely completion of a majority of MMLS tasks. Extensive evaluation results with NS-3 show that the overall live latency is kept within an acceptable range, and the probability of rebuffering events is significantly reduced compared to the benchmark schemes.
编者译:为解决大规模移动直播中的高延迟与卡顿问题,本文提出了一种名为GOMOS的基于图神经网络的网络与媒体协同优化框架。不同于传统单视角优化,GOMOS充分考虑网络与媒体间的相互作用。该框架通过三阶段实现协同:首先,利用双向链路探测算法,实时收集并分发有线网络的状态信息;其次,利用图神经网络模型,综合网络与媒体特征,精准预测海量直播流在有线链路上的传输延迟;最后,将无线资源分配问题建模为跨层优化问题,动态调整资源以确保直播任务及时完成。在NS-3平台上的大量实验表明,GOMOS能在将整体延迟控制在可接受范围内的同时,显著降低直播卡顿的发生概率,效果优于基准方案。
题目:基于重排序和后检索反思的教育大模型问答增强方法
作者:孙浩然、王志豪、吴一帆、高晓影、向阳
来源:大数据,2025,11(05):4-17.
摘要:介绍了大语言模型在教育领域的应用,特别是在计算机教育中的知识问答系统。大模型通过预训练获得推理、长文本建模和文本生成能力,但在特定领域如计算机知识问答中存在知识不具体和“幻觉”问题。为解决这些问题,提出了有监督微调、CoVe框架和检索增强生成(RAG)技术。然而,传统RAG技术存在检索能力不足和召回无关材料的问题。本文提出了一种基于重排序和后检索反思的教育大模型问答增强方法,通过构建中文和英文外置知识数据库,利用RAG技术改善大模型的知识性缺陷。该方法在RAG过程中引入基于交叉编码器的重排序模型,计算检索结果与用户查询的相关性得分,过滤无关材料,使辅助知识更准确。同时,引入后检索反思环节,利用初轮回答结果进行检索和重排序查找辅助知识,用于生成模型自我纠错,提升回答准确度。实验结果表明,该方法适用于不同生成模型,提升了大模型在计算机知识问答中的正确率,缓解了“幻觉”现象。
题目:Mathematical Modeling and Deployment Optimization: Intelligent Reflecting Surface-Aided Communications Under Partial Blockages(数学建模与部署优化:部分阻塞环境下智能反射面辅助通信)
作者:H. Hashida, Y. Kawamoto and N. Kato
来源:IEEE Transactions on Cognitive Communications and Networking(认知通信与网络汇刊), vol. 11, no. 5, pp. 3306-3316.
摘要:Intelligent reflecting surfaces(IRS)can eliminate dead zones by providing alternative paths for radio signal propagation. However, in scenarios with many moving obstacles, the IRS aperture can be partially blocked, which can decrease the effective area and, consequently, reduce its gain. Although enlarging the IRS aperture can mitigate the effects of these obstructions, large IRSs can introduce inefficiencies because of location and cost constraints. Therefore, in this study, we propose a novel method for optimally designing IRS installation locations by considering spatial characteristics, such as the size and density of obstacles. We perform numerical analysis to highlight the importance of considering moving obstacles in developing an effective IRS deployment strategy. This study provides insights for establishing communication zones in environments with numerous obstructions, including smart factories, offices, and warehouses.
编者译:智能反射面(IRS)能够为无线电信号传播提供替代路径,从而消除通信盲区。然而,在移动障碍物较多的情况下,智能反射面孔径可能被部分阻挡,导致有效反射面积减小,进而降低其增益。虽然增大反射面孔径可以缓解此类阻挡造成的影响,但受安装位置和成本因素制约,大型反射面往往会带来效率低下的问题。为此,本研究提出一种创新方法,通过综合考虑障碍物尺寸与密度等空间特征,实现反射面安装位置的优化设计。通过数值分析,本文阐明了在制定有效反射面部署策略时考虑移动障碍物的重要性。本研究为在智能工厂、办公室及仓库等多障碍物环境中建立可靠通信区域提供了重要参考。
题目:生成式AI搜索引擎人机结合的选题思路拓展研究
作者:王树义、曾雯、戚淇、许隆鑫、岳芳
来源:图书情报知识, 2025, 42(4): 113-125.
摘要:针对研究现状提出基于生成式AI搜索引擎的选题工作流,通过实验和访谈的方法对选题过程进行验证。结果显示并依据结果讨论“人在环中”在选题过程中的优势。所有被试对选题拓展效果的满意度评分均在90分及以上(满分100分),生成式AI搜索引擎可以有效辅助科研选题,但需要采取人机结合的方式才能发挥最大效用;“人在环中”可以充分发挥AI的优势,提升选题质量和效率,推动科学研究向更高层次发展。基于生成式AI搜索引擎与人机结合的选题思路拓展方法,为科研工作者提供了一种全新的选题思路和方式,有助于缓解信息过载和研究选题困难的问题,为科研工作者提供了理论指导。
题目:市域数据要素市场构建路径探索:无锡创新实践的经验与启示
作者:胡逸、颜春水、章瑜桢、蒋子海
来源:大数据,2025,11(05):117-129.
摘要:党的二十届三中全会强调培育全国一体化技术和数据市场,加强数据要素市场建设对促进数据流通利用、释放数据价值至关重要。目前全球数据流通利用处于起步阶段,欧盟、美国和中国分别探索不同政策和治理模式。学术界也在研究数据要素市场建设,提出联动路径和评估体系。无锡在市域层面提出数据要素市场建设路径,介绍“中国数码头”框架,为数据驱动经济转型提供有益启示。
