Network-splitter:一种基于重叠社区的网络特征提取算法及其在链路预测中的应用

Network-splitter:a network feature extraction algorithm based on overlapping community and its application in link prediction

链路预测任务是根据已知的网络结构和节点属性等信息来预测网络中产生新链路的可能性.它是网络科学中的一个基础性问题,具有重要的理论研究和实际应用价值.近年来,网络表示学习领域的学者利用深度学习提取网络复杂特征,大幅度提高了链路预测效果.实际网络中节点具有局部聚类现象,然而,当前的网络表示学习侧重于提取网络全局特征,忽略了局部信息特征.针对这个问题,我们提出了能够学习网络中节点在不同社区中局部特征表示的模型network-splitter.该模型利用重叠社区思想,在每个社区中创建节点的一个角色副本,并学习该角色副本的特征表示.最后将节点在不同社区中对应的角色副本信息通过神经网络综合,得到的综合向量包含网络全局特征和节点局部特征,并可应用到链路预测任务中.本文的实验结果表明,network-splitter模型与最新的网络学习表示方法相比具有很强的竞争力.

Link prediction is the task of forecasting the possibility of generating new links in a network based on network structure and node attributes.It is a basic problem in network science and valuable both in theory and practice.In recent years,deep learning methods are widely used for network representation to extract complex network features,which greatly improves the results of link prediction.The nodes in real-world networks have the phenomenon of local clustering;however,the current network representation learning methods focus on extracting the global features of the network,ignoring the local information features.In order to solve this problem,we propose a network-splitter model,which can learn the local feature representation of nodes in different communities.The model uses the idea of overlapping community to create a role copy of a node in each community and learns the feature representation of the role copy.Finally,the role copy information of the node in different communities is synthesized through the neural network.Using this method,both the global feature of the network and the local features of the nodes are summarized into the network representation,and then are applied to the link prediction.The experimental results show that the network-splitter model has strong competitiveness compared with the latest network representation learning methods.