In this paper, we present a non-parametric overlapping community detection method based on the affiliation graph model using an Indian Buffet Process to determine the number of communities. We compare this model with a full stochastic blockmodel using the same prior, as well as two other models, a blockmodel and a community model, that employ non-parametric priors based on a Gamma Process. We ask two questions; firstly, whether community models are sufficient to model the overlapping structure of real networks, without resorting to blockmodels that entail significantly more parameters; secondly, which is the better non-parametric approach of the two analysed? Measuring performance in terms of predicting missing links, we find that all models obtain similar performance, but in general, the Indian Buffet Process prior results in simpler models, with fewer blocks or communities. We argue that, when obtaining the latent structure is the purpose of the analysis, the simpler affiliation graph model, with Indian Buffet Process is preferred.
CITATION STYLE
Laitonjam, N., & Hurley, N. (2019). Non-parametric overlapping community detection. In Springer Proceedings in Mathematics and Statistics (Vol. 296, pp. 23–34). Springer. https://doi.org/10.1007/978-3-030-30611-3_3
Mendeley helps you to discover research relevant for your work.