Traditional link prediction methods of social network are vulnerable to the influence of network structure and have poor generalization, and only on a small number of networks and evaluation indicators. To improve the stability and accuracy of link prediction, this paper assembles 15 similarity indexes, introduces the idea of stacking into the link prediction of complex networks, and presents a link prediction method (Logistic-regression LightGBM Stacking Link Prediction, LLSLP). Firstly, social network link prediction is regarded as a binary classification problem. Secondly, the hyper parameters of the basic model are determined by using cross-validation and grid searching; thirdly, Logistic-regression and LightGBM are integrated by stacked generalization; Finally, take 10 different networks as practical examples. The feasibility and effectiveness of the proposed method are verified by comparing 7 evaluation indicators. The experimental results show that: the proposed method is not only more than 98.71% higher than the traditional CN (Common Neighbor) and other models are 10.52% higher on average. In addition, compared with the traditional 15 link prediction algorithms, $F1- score$ value and $MCC$ (Matthews Correlation Coefficient) value is increased by 3.2% ~ 9.7% and 5.9% ~ 14% respectively. The proposed method has good accuracy and generalization. It can also be applied to recommendation system.

中文翻译：

---

一种基于堆叠泛化的社交网络链接预测方法

传统的社交网络链接预测方法易受网络结构的影响，泛化性差，且仅适用于少数网络和评价指标。为了提高链路预测的稳定性和准确性，本文汇集了15个相似度指标，将stacking的思想引入到复杂网络的链路预测中，提出了一种链路预测方法（Logistic-regression LightGBM Stacking Link Prediction，LLSLP）。首先，社交网络链接预测被视为一个二元分类问题。其次，通过交叉验证和网格搜索确定基本模型的超参数；第三，Logistic-regression和LightGBM通过stacked generalization进行整合；最后，以 10 个不同的网络为例。通过对比7个评价指标，验证了所提方法的可行性和有效性。实验结果表明：所提出的方法不仅比传统的CN（Common Neighbor）高出98.71%以上，其他模型平均高出10.52%。此外，与传统的15个链路预测算法相比，$F1- score$值和$MCC$（Matthews Correlation Coefficient）值分别提高了3.2%~9.7%和5.9%~14%。该方法具有良好的准确性和泛化性。它也可以应用于推荐系统。此外，与传统的15个链路预测算法相比，$F1- score$值和$MCC$（Matthews Correlation Coefficient）值分别提高了3.2%~9.7%和5.9%~14%。该方法具有良好的准确性和泛化性。它也可以应用于推荐系统。此外，与传统的15个链路预测算法相比，$F1- score$值和$MCC$（Matthews Correlation Coefficient）值分别提高了3.2%~9.7%和5.9%~14%。该方法具有良好的准确性和泛化性。它也可以应用于推荐系统。