一种基于高阶混合投影估计的网络嵌入方法

特征提取对于网络分析任务而言是至关重要的,而网络嵌入学习的目的则是根据网络的结构和语义信息自动化构建节点或边的特征。现有的方法将网络嵌入分解为网络数据挖掘和数据降维两个独立的过程,因而无法很好地在潜在空间中对节点的分布进行建模描述。因此,提出了一种基于高阶混合投影估计的网络嵌入方法,该方法借鉴谱分解的思想,利用线性投影算子将网络从高维结构空间映射至低维特征空间,然后利用混合概率模型对节点的分布进行建模以维持网络的社区结构性质。此外,该方法还融入了局部节点相似性来防止发生过拟合现象。最后,为了验证该方法的有效性和鲁棒性,在四个真实的网络数据集之上和现有的网络嵌入算法进行了对比实验,在链路预测任务中,该方法分别将Micro-F1和Macro-F1指标的基准线平均提升了3.97%和2.23%,在节点分类任务中,该方法将AUC值的基准线平均提升了10.43%。

A Network Embedding Method Based on High-order Hybrid Projection Estimation

Feature extraction is crucial for network analysis,while the purpose of network embedding learning is to automatically generate discriminable features for each nodes or edges based on the network structure and semantic information.The existing approaches decouple network mining and data reduction into two independent stages,which fails to model the distribution of the nodes in the latent space.Therefore,a novel network embedding method based on high-order hybrid projection estimation is proposed.Inspired by the idea of spectral-based embedding,this method utilizes a linear orthogonal projection operator to map the network from the high-dimension structural space to the latent low-dimension space.Besides,the distribution of nodes is modeled by a mixture probability model to maintain the community structure of the network.Moreover,the local proximity between nodes is considered as a constraint to avoid over-fitting.To evaluate the effectiveness and robustness of the proposed method,the comparison experiment is carried out on four real network data sets with other network embedding methods.In the link prediction task,the model achieves an average gain of 3.97% for Micro-F1 and 2.23% in Macro-F1.In the node classification task,the model improves the best baseline on average 10.43% in AUC score.