Abstract
Community structures are common features of many real-world networks, and community detection is necessary to understand how these networks are organized. Various approaches have been devised for community detection, with each providing varying degrees of both accuracy and structural understanding. One of them, the Label Propagation Algorithm, is so common because it is simple and computationally cheap. Nevertheless, it does not usually reach great modularity and yields inaccurate community counts and structures in real-world networks. This is mostly due to its naive criteria of selecting the neighbor nodes when it comes to label propagation. To tackle the issue, we developed an adjusted algorithm, which we call Embedding-based Label Propagation (ELP), a hybrid between LPA and node embedding that allows us to combine both local connectivity and global structural data. ELP update step takes into consideration not only the local neighborhood, as in conventional LPA, but also embedding-based similarities to inform more productive neighbor selection. We tested ELP on popular benchmark datasets such as Karate Club, Dolphins, Football, Polbooks, and LFR synthetic networks and compared its results with LPA and other well-established algorithms. The empirical findings show that ELP can always perform better in modularity, NMI and NF1 scores, but it is also scalable to large and complex networks. These results can be used to identify ELP as an effective and powerful method of community-finding in real and artificial-world scenarios.