Joint Learning of Hierarchical Community Structure and Node Representations: An Unsupervised Approach

Graph representation learning has demonstrated improved performance in tasks such as link prediction and node classification across a range of domains. Research shown that many natural graphs can be organized hierarchical communities, leading to approaches use these communities improve the quality representations. However, do not take advantage learned representations also discovered establish an iterative joint optimization community discovery. In this work, we present Mazi, algorithm jointly learns structure graph unsupervised fashion. To account for representations, Mazi generates at each level hierarchy, utilizes them influence original graph. Further, are by simultaneously maximizing modularity metric minimizing distance between its community. Using multi-label tasks, evaluate our method on variety synthetic real-world demonstrate outperforms other non-hierarchical methods.