Research On The Influence Of Network Structure Characteristics On Link Prediction Algorithms Based On Meta-analysis

ObjectiveWith the rapid development of Internet technology since the 21st century,a large amount of data has gradually increased exponentially,and the difficulty of data acquisition and processing has gradually decreased.Common complex systems such as power network,transportation network,social network and so on have begun to be modeled as complex networks for in-depth research.Especially in the past decade,complex network technology has been gradually promoted and applied to many disciplinary fields,becoming the topic of preface of various disciplinary fields.The formation and evolution of network,the relationship between network topology and function,and the characteristics of network topology structure have also become the key issues to be discussed in the theory and application of complex network.As one of the core branches of complex network technology,link prediction can predict the probability of future link between disconnected nodes based on the existing network topology,which is of great value to reveal the evolution law of complex network.From the perspective of the relationship between network topology and functionality,we can understand that the effectiveness of link prediction algorithms has a wide range of performance in a variety of networks.Based on these differences in performance,this study intends to integrate data from several existing link prediction experiments to build a meta-analysis statistical test model to comprehensively explore the issue of how network topology affects the effectiveness of link prediction algorithms.MethodsIn this study,we retrieved relevant journal papers on the topic of link prediction from10 databases,including CNKI,Web of Science,EBSCO,EI,IEEE,Pub Med,Pro Quest,Scoups,Science Direct,and Springer.A Three-level Meta-analysis Model is built to explore whether there is a difference in the effectiveness of link prediction algorithms among complex networks with different topologies.The main moderating variables affecting the effectiveness of link prediction algorithms are examined through a multifactor meta-regression framework of three-level meta-analysis,and based on this,a subgroup analysis of the moderating variables is conducted using a Bayesian network meta-analysis model to explore the influence of the moderating variables on the effectiveness of link prediction algorithms in conjunction with the cumulative probability ranking and SUCRA values.ResultsThrough literature retrieval and screening,a total of 22 complex networks and 26 link prediction algorithms from 5 papers were included in the study,including 278 effect sizes.Meta-analysis statistical test results showed that the overall average effect size of three-level meta-analysis was MD=1.18（95%CI:[1.00,1.37]）.Among the seven network topology features included in the data screening conditions,Three parameters showed significant effect(P_val<0.05),namely network density,network average degree and aggregation coefficient.Subgroup analysis results show that high average degree has a positive effect on local information based link prediction algorithms HDI（SUCRA=0.8825）,Salton（SUCRA=0.8286）,HPI（SUCRA=0.6705）and RA（SUCRA=0.6635）.Network density is the main parameter affecting PA index,and it shows great difference between high level network density（SUCRA=0.6633）and low level network density（SUCRA=0.0241）.In sparse networks,the SUCRA values of SRW,LRW,LHN-II and MFI are all greater than 0.80,showing strong positive predictive power.Katz index shows strong positive effect under the condition of low average degree and low network density,SUCRA value is 0.8089 and 0.8632,respectively.Aggregation coefficient has little influence on Katz algorithm,however,the network with high aggregation coefficient（SUCRA=0.9063）performed better than the network with low aggregation coefficient（SUCRA=0.7523）.Low average degree and high aggregation coefficient had moderate positive influence on LP index,and their SUCRA value was0.6521 and 0.7654,respectively.ConclusionsThe results of three-level meta-analysis show that the effectiveness of the link prediction algorithm varies greatly among different networks,and network density,network averageness and aggregation coefficient are the main factors affecting the difference.The results of subgroup analysis show that in dense networks,the link prediction algorithm needs to consider the local information more,and the link prediction algorithm based on local information is more efficient.In contrast,algorithms based on quasi-local information and global information take more into account the overall structure of the network and play a more significant role in sparse networks.