基于节点度异质性惩罚的链路预测方法

Link Prediction Method Based on Node Degree Heterogeneity Penalization

针对大部分现存的链路预测方法仅关注规则网络以及偏好连接现象而导致在稀疏网络获得低质量性能,提出一种节点度异质性惩罚的链路预测框架(NDHP),该框架最优预测准确度与网络拓扑特征有密切关联。首先,计算整个网络节点度获得所有节点对的度异质性相似度;其次,采用惩罚节点度较大机制去惩罚度异质性权重较大的节点抑制节点间差异;最后,通过可调参数将平均节点聚类系数和平均最短路径分别和基于度异质性惩罚框架相关联,获取网络结构信息来弥补网络稀疏信息不足,并提出基于节点度异质性惩罚的平均聚类系数指标(NDHP_AC)和基于节点度异质性惩罚的平均距离指标(NDHP_AD)。此外,在8个真实无向无权网络上与最近代表性的方法相比较,所提两个指标在预测缺失链接和鲁棒性两方面性能优于基准指标。尤其在高度稀疏网络中,所提指标的AUC和AUPR分别最大提高了15.3%和8.6%。

Most of existing link prediction methods only focus on regular networks and preferential attachment phenomenon,which results in low quality performance in highly sparse networks.We propose a link prediction framework based on node degree heterogeneity penalty.The optimal prediction accuracy of the framework is closely related to network topology characteristics.Firstly,the node degree of the whole network is calculated to obtain the degree heterogeneity similarity of the predicted node pairs.Secondly,the mechanism of punishing nodes with higher degree of heterogeneity is used to suppress the differences between nodes.Finally,the average node clustering coefficient and the average shortest path are associated with the degree heterogeneity based punishment framework respectively by the adjustable parameters to obtain the network structure information to make up for the lack of sparse network information,and the Node Degree Heterogeneity Penalization via Average Clustering coefficient(NDHP_AC)and the Node Degree Heterogeneity Penalization via Average Distance(NDHP_AD)are proposed.In addition,compared with the most recent representative method,the performance of the proposed two indicators is better than the benchmark indicators in predicting missing links and robustness on eight real undirected and unweighted networks.Especially in highly sparse networks,the proposed method improves the maximum AUC and AUPR by 15.3%and 8.6%,respectively.