中文摘要：

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.

英文摘要：

Network feature learning can obtain the low-dimensional representations of network by analyzing the relationships（structures or attributes）between nodes.However,there are many nodes embedding methods,which based on assumptions of static and small-scale（such as language networks）,are unable to adapt to social networks,because social networks have their specific properties such as dynamic and large-scale.Based on current researches,this paper propose a damping based positive and negative sampling model for learning nodes embedding of social networks.By sampling nodes at different levels with damping,at the same time design an incremental learning method for newly added nodes,which makes it possible for learning nodes features extracting during the dynamic changing process,thus to learning a better representations of social networks.Finally,we select three large-scale,dynamic and real-life social networks for dynamic link prediction task.The results show that,compared with DeepWalk and LINE methods,DNPS have achieved greater performance in prediction accuracy and time efficiency.The learned node vectors by DNPS model can be used in many subfields of social network research.For example,we can use it for large-scale dynamic social community discovery,user recommendation,and user labeling.