基于循环神经网络的在线动态网络嵌入算法

AN ONLINE DYNAMIC NETWORK EMBEDDING ALGORITHM BASED ON RECURRENT NEURAL NETWORK

网络嵌入旨在为网络中的节点学习低维的向量表示。大部分现有算法只适用于静态网络,然而对于现实世界不断增长的网络需要重新训练,降低了方法的可用性。对此提出循环网络嵌入(Recurrent Neural Network Embedding,RNNE)来处理在线动态网络。为了解决网络规模可能频繁改变的问题,RNNE在网络中添加了独立的虚拟点保持网络在不同时间点规模的统一。同时,RNNE在嵌入时兼顾了网络的静态和动态特征,一方面通过节点间的连边和邻居的相似度来保持网络的局部和全局结构,另一方面通过传递先前时刻的嵌入信息来减少噪音的影响。RNNE在5个数据集上与其他几个最新的算法进行了测试和比较,结果表明RNNE相比于这些算法在重构、节点分类和链路预测上具有更大的优势。

Network embedding means learning a low-dimensional representation for every node in the network.Most of the existing algorithms can only deal with static networks,but the growing network in the real world requires retraining,which reduces the usability of the method.To solve this problem,this paper proposes a recurrent neural network embedding(RNNE)algorithm to deal with online dynamic network.In order to solve the problem that the network scale might change frequently,RNNE added independent virtual node to the network to keep the scale of the network unified at different time.RNNE took into account the static and dynamic characteristics of the network when embedding.On the one hand,it maintained the local and global structure of the network through the edge connection between nodes and the similarity of neighbors.On the other hand,it reduced the impact of noise by transmitting the embedded information at the previous time.RNNE was tested and compared with several new algorithms on five data sets.The results show that RNNE has greater advantages in reconstruction,node classification and link prediction than these algorithms.