图池化作为图神经网络中重要的组件,在获取图的多粒度信息的过程中扮演了重要角色。而当前的图池化操作均以平等地位看待数据点,普遍未考虑利用邻域内数据之间的偏序关系,从而造成图结构信息破坏。针对此问题,本文提出一种基于偏序关系的多视图多粒度图表示学习框架(multi-view and multi-granularity graph representation learning based on partial order relationships,MVMGr-PO),它通过从节点特征视图、图结构视图以及全局视图对节点进行综合评分,进而基于节点之间的偏序关系进行下采样操作。相比于其他图表示学习方法,MVMGr-PO可以有效地提取多粒度图结构信息,从而可以更全面地表征图的内在结构和属性。此外,MVMGr-PO可以集成多种图神经网络架构,包括GCN(graph convolutional network)、GAT(graph attention network)以及GraphSAGE(graph sample and aggregate)等。通过在6个数据集上进行实验评估,与现有基线模型相比,MVMGr-PO在分类准确率上有明显提升。
在互联网时代,网络营销迅速发展,社交媒体的作用已超越单纯的社交分享,电子商务与新媒体的结合日趋成熟,衍生出多种细分模式。作为中国影响力较大的社交媒体平台,微博近年来吸引了大量关于消费者行为的学术研究,重点关注内容营销、互动营销、品牌推广等因素。相比之下,国外以“湿营销理论”为代表的研究起步较早,而国内研究多聚焦于口碑影响和营销价值等方面。本研究以大学生为调查对象,提取了在线客户服务、内容营销、互动营销和品牌推广四种微博营销形式。通过问卷收集数据后,结合案例分析发现,企业在微博营销中应重视内容营销以增强用户粘性,合理利用明星效应以提高广告效果,并完善售后服务提升客户心理价值。多元营销形式的组合有助于企业在微博平台上吸引关注、影响消费者购买决策并实现收益最大化。In the Internet era, online marketing has developed rapidly, the role of social media has gone beyond mere social sharing, and the combination of e-commerce and new media has become increasingly mature, giving rise to a variety of segmentation models. As an influential social media platform in China, Weibo has attracted a lot of academic research on consumer behaviour in recent years, focusing on content marketing, interactive marketing, brand promotion and other factors. In contrast, foreign research, represented by the “wet marketing theory”, started earlier, while domestic research focuses on word-of-mouth influence and marketing value. This study takes college students as the survey object and extracts four Weibo marketing forms: online customer service, content marketing, interactive marketing and brand promotion. After collecting data through questionnaires and combining with case studies, it is found that enterprises should pay attention to content marketing in Weibo marketing to enhance user stickiness, make reasonable use of celebrity effect to improve advertising effect, and improve after-sales service to enhance customers’ psychological value. The combination of multiple marketing forms can help enterprises attract attention, influence consumers’ purchasing decisions and maximise revenue on the Weibo platform.
半监督学习由于能够充分利用未标记数据而广受关注,其中图半监督学习方法具有表示直观和概念清晰的优点。然而,基于图的半监督学习方法大多需要迭代优化,且由于初始标记点的选取变化,会导致预测准确性不稳定。为了解决这一问题,提出了一种基于高密度近邻和确定性标记(high density nearest neighbors and determinate labeling,HDN-DL)的半监督分类方法,利用数据中的潜在结构和信息,选择影响力较高的节点作为标签传播的起点,通过密度峰值聚类算法(density peak clustering,DPC)得到初始的无监督聚类图结构后,将该图断开得到引领森林,再根据每个样本点的所在层次计算其高密度近邻及其相对距离,以此来综合考虑多个属性以判定当前样本的标签,避免级联误分。标签传播的过程无需迭代,复杂度为O(n)。在多个数据集上进行的实验验证了该方法的有效性和稳定性。
暂无评论