基于卷积神经网络的链接表示及预测方法

针对节点全局表示和链接局部拓扑关系，提出链接序列化表示及卷积神经网络（CNN）提取序列特征的链接预测方法.研究节点间的局部拓扑及共邻关系，基于共邻紧密度构建链接局部拓扑的有序节点序列，并用node2vec节点向量表达生成潜在链接的矩阵表示；基于CNN建立链接预测的分类模型，采用CNN可变滤波器窗口卷积运算提取序列中共邻与节点对的多层隐含关系，分类训练实现链接的有效预测.在4种大规模网络数据集上的实验结果表明，相比已有方法，该方法的AUC值有显著提高，最高达12.4%，稳定性及普适性较强，解决了传统方法对大规模稀疏网络的预测准确率下降问题.
Abstract: A novel link prediction method of link serialized representation and convolution neural network (CNN) extracting sequence features was proposed in view of the global node representation and the link's local topology. The local topological and common neighbors' relationship between node pairs were studied. Based on common neighbors' density, the ordered node sequences of links' local topology were constructed, and the matrix representation of potential links were generated by node2vec to express the nodes in the sequences vectorically. The classification model of link prediction was established based on CNN. The variable-size filter windows in CNN were used for convolution operation to extract the multilayer implicit relation features among common neighbors and node pairs in the sequences. After the training and classification process, potential links were effectively predicted. The experimental results on four large scale network datasets show that this method has significantly improved AUC value of up to 12.4% compared with the existing methods, and has strong stability and universality, which also solves the problem that traditional methods' accuracy for link prediction decreases with the large-scale sparse network.