为使链路预测应用于大型复杂网络,设计并实现了一种基于MapReduce 计算模型的并行链路预测算法,包含了9 种基于局部信息的相似性指标,在稀疏网络上的时间复杂度为O(N).首先,在公共数据集上验证了并行算法的有效性,随着抽取因子的增加,召回率升高而准确率下降.在不同类型的10 个大规模复杂网络数据集上的实验结果表明,基于MapReduce 计算模型的并行链路预测算法比传统算法具有更高的效率,算法的运行时间随着并行程度的增加而下降.提出并证明了AUC(area under a receiver operating characteristic curve)评价指标的上下界,实验表明,上下界的中值和实际AUC 值很接近,并且AUC 评价指标侧重于预测分数值是否为0 而不是分数值的大小.在网络拓扑性质中,平均聚集系数对AUC 值的影响最大,并且AUC 值随着网络平均聚集系数的增加而提高.

To apply link prediction methods into large-scale complex network, this paper designs and implements a parallel link prediction algorithm based on MapReduce, which includes nine similarity Indices via local information. The parallel link prediction algorithm has a time complexity of O(N) in sparse networks. First, the paper verifies the validity of the algorithm on public datasets, increase in the extraction factor, recall ascends, and precision descends. The experimental results on ten large-scale datasets of variety network types show that the parallel link prediction algorithm is more effective than traditional ones, and its running time decreases with more compute units. The upper and lower bounds of AUC (area under a receiver operating characteristic curve) are proposed. The experimental results show the median of the upper and lower bounds are close to the real value of AUC, which focuses on whether prediction score is zero rather than the actual score value. The network average clustering coefficient has the greatest impact on AUC among most topological features and AUC rises as the network average clustering coefficient increases.

国家自然科学基金(90924029, 60905025, 61074128)