网络结构对链路预测算法的影响研究——基于元分析视角

Network Structure’s Impacts on Link Prediction Algorithm from Meta-Analysis Perspective

【目的】在不同结构的网络中,各链路预测算法的预测效果存在不同程度的差异,通过对多项研究的数据进行统计分析,可以系统挖掘网络结构特征中影响链路预测结果的主要参数。【方法】选取国内外关于链路预测的相关实证研究,最终纳入5篇文献、22个网络、26种算法和278项研究,利用三水平元分析和贝叶斯网络元分析方法探讨网络结构中影响链路预测结果的主要因素及其对各算法预测结果的影响。【结果】纳入研究的算法总体预测的效应量MD=1.1832(95%CI:(1.0005,1.3659)),网络密度、平均度和聚集系数是影响各算法预测效果的主要因素(Pval<0.05)。亚组分析结果表明:Katz、LHN-II、MFI、LRW、SRW等基于全局信息和准局部信息的链路预测算法在稀疏网络性能更佳,SUCRA值均大于0.5,在稠密网络中网络密度、网络平均度和聚集系数对各类算法的影响差异较大。【局限】仅从统计学的角度进行分析,并未纳入大规模的文献数据进行进一步的实证分析,结果还较为粗糙。【结论】本研究将元分析的概念引入复杂网络领域中,丰富了对网络结构与链路预测算法关系探讨的方法与视角,为未来网络结构对链路预测算法影响的相关研究提供新的思路。

[Objective]This paper tries to identify the main influencing parameters of the link prediction algorithms with the help of network structures and data from multiple studies.[Methods]We retrieved empirical research on link prediction from China and abroad,which include 5 papers,22 networks,26 algorithms and 278 studies.We used three-level meta-analysis and Bayesian network meta-analysis to explore the network structures and their impacts on algorithms’performance.[Results]The algorithms included in our study generally had a good predictive effect MD=1.1832(95%CI:(1.0005,1.3659)).The network density,average degree and clustering coefficient are the main factors affecting the prediction results(Pval<0.05).Katz,LHN-II,MFI,LRW,and SRW algorithms yielded better results with sparse networks and their SUCRA values were greater than 0.5.[Limitations]Our research does not include empirical analysis with large-scale data.[Conclusions]With the help of meta-analysis,our study explores the development directions for the link prediction algorithms.