表征学习驱动的多重网络图采样

已有的图采样方法侧重于单图采样，关注如何在一张图上通过采样保留其特定的拓扑结构特征。随着数据采集能力的提升，多重网络图在实际应用中越来越普遍，即相同的节点集在不同场景中具有不同的网络关系。针对传统图采样方法无法兼顾多重网络图结构特征的问题，提出了表征学习驱动的多重网络图采样算法。首先，设计融合多重网络图结构特征的图表征学习方法，将节点投影至二维的表征学习空间；其次，利用改进的自适应蓝噪声采样算法，考虑节点密度和网络连通性，从表征学习空间筛选节点，以保持其多重网络结构特征及图上下文结构特征。进而开发了一套多重网络图采样可视分析系统，支持用户交互式地探索多重网络图采样，并与已有采样算法进行对比。案例分析和评估实验证明了本文算法在多重网络图采样中的有效性。

Representation learning driven multiple graph sampling

The existing graph sampling techniques pay attention mainly to single graph sampling, focusing on how to preserve the specific topological features of a graph by sampling, such as node degree, cluster coef?cient, connectivity. With the improvement of data acquisition capability, multiple graphs, namely a set of nodes exhibits different relationships in different scenarios, have become quietly ubiquitous in the real world. To address this problem, a multiple graph sampling driven by representation learning is proposed. First, a graph representation learning method is designed to fuse the structural features of multiple graphs, through which the nodes are projected into a two-dimensional representation learning space. Then, considering node density and network connectivity, an improved the adaptive blue noise sampling algorithm is employed to select nodes from the representation learning space meanwhile simultaneously preserving the contextual structure features of multiple graphs. Furthermore, an interactive visual analytics system is developed allowing users to explore and analyze multiple graph sampling, and visually compare results with various sampling strategies. Case studies and the experimental results based on two real-world datasets have demonstrated the effectiveness of the proposed method in sampling multiple graphs.