Detecting Overlapping Communities in Knowledge Graphs: A Density Optimization Based Approach

Detecting overlapping communities in knowledge graphs is considered a problem of fundamental importance, since the growing trend of interdisciplinary makes it common for a piece of knowledge belong to different realms. Among all algorithms detecting overlapping communities, Speaker-listener Label Propagation Algorithm (SPLA) represents one of the state-of-the-art approaches due to its high accuracy. However, in the face of overlapping communities with instability and unbalanced partitions, the performance of SLPA drastically degrades. To fill such a gap, we propose a novel community detection algorithm based on density optimization. The proposed algorithm leverages Jaccard similarity coefficient to quantify the similarity between two nodes, and then propagate the label of a node to its neighboring node with the highest similarity. In this way, the initial community is obtained. Due to the existence of an oversized community in the initial community, we argue that a good community structure should have a higher density within the community than outside the community. Therefore, the initial community should be divided again. If the density of the new community is larger than that of the original community, the node's label information is updated; otherwise not. Finally, extensive experiments are carried out on both artificial networks and real networks. The results show that the proposed approach can achieve higher NMI index and overlapping modularity, hence outperforming existing methods.