DNPS:基于阻尼采样的大规模动态社会网络结构特征表示学习

发布时间：2018-06-14 00:11

本文选题：社会网络 + 节点嵌入　；参考：《计算机学报》2017年04期

【摘要】：网络特征表示学习通过对网络节点之间的关系(结构或属性)进行分析,得出网络特征的低维度表达.现有的针对网络特征学习的方法多基于静态和小规模的假设(如静态的语言网络),并没有针对社会网络的特有属性进行修正学习,因此,现有的学习方法无法适应当前社会网络所具备的动态性、大规模甚至超大规模等特性.该文在已有研究基础上,提出了基于动态阻尼正负采样的社会网络结构特征嵌入模型(Damping Based Negative-Positive Sampling of Social Network Embedding,DNPS).通过对不同阶层的网络节点关系进行正负阻尼采样,同时构建针对新增节点的动态特征学习方法,使得模型对于大规模社会网络在动态变化过程中的结构特征的提取变得可行,以此获得的节点特征表达具备更好的动态鲁棒性.通过选取3个大规模的动态社会网络的真实数据集和在社会网络的动态链接预测问题的实验中发现:DNPS相对于基准模型(DeepWalk/LINE)在预测准确率以及时间效率上都取得了较大的性能提升.同时,DNPS的学习结果还可以被应用于社会网络的相关研究子领域.例如,在大规模以及动态性的环境下,研究大规模动态社区发现、社会网络用户推荐、标记分类等问题.
[Abstract]:By analyzing the relationship between network nodes (structure or attribute), the low dimension expression of network features is obtained. Most of the existing methods for learning network features are based on static and small-scale assumptions (such as static language networks), and do not modify learning for the specific attributes of social networks, so, The existing learning methods can not adapt to the dynamic, large-scale or even super-scale characteristics of current social networks. In this paper, a dynamic damped positive and negative sampling model based on dynamic damped positive and negative sampling is proposed to embed the structural features of social network, which is called Damping based Negative-Positive sampling of Social Network embedding (DNPSN). By sampling positive and negative damping of network nodes at different levels and constructing a dynamic feature learning method for new nodes, it is feasible to extract the structural features of large-scale social networks in the process of dynamic change. The node feature representation obtained by this method has better dynamic robustness. Through selecting the real data sets of three large-scale dynamic social networks and the experiments on the dynamic link prediction problem of social networks, it is found that the performance of the prediction accuracy and time efficiency of the "DNPS" is higher than that of the benchmark model (DeepWalk-LINE). At the same time, the learning results of DNPS can also be applied to the relevant research subfields of social networks. For example, in the large-scale and dynamic environment, the problems of large-scale dynamic community discovery, social network user recommendation, label classification and so on are studied.
【作者单位】：中国人民大学信息学院;中国人民大学信息资源管理学院;
【基金】：国家自然科学基金(71271211,71531012) 北京市自然科学基金(4172032) 中国人民大学科学研究基金(10XNI029) 中国人民大学2016年度拔尖创新人才培育资助计划成果资助~~
【分类号】：TP181

【相似文献】