Abstract:
Artificial intelligence (AI) is comparable to human intelligence in many cases. Especially in the case of graph-structured data, there are many AI algorithms and machine learning models which not only can understand these graph-structured data successfully but are also able to predict the future with high precision. Many researchers developed machine learning models to be applied to the graph-structured data. The most successful machine models acting on the graphs are the graph convolutional networks (GCNs). Some research has considered applying GCNs to tackle the overlapping community detection problem. Indeed, overlapping community detection is a key problem in graph-structured data mining. However, the existing GCN-based approach considered a shallow (2-layered) GCN which has an essential limitation in the detection of communities with larger sizes. A viable solution can be the use of deep GCNs. However, training deeper GCNs is notoriously difficult due to several key issues like over-smoothing and vanishing/exploding gradient problems. Some researchers have attempted to resolve these issues in the case of regular graphs generated from point clouds. Unfortunately, it is still challenging how to incorporate deep GCNs in the case of general irregular graphs. In this study, we have resolved the issues of deep GCNs in the case of irregular graphs in an overlapping community detection framework. In particular, we design a deep dynamic residual GCN (DynaResGCN) based on our dynamic dilated aggregation mechanisms in general irregular graphs. Additionally, we introduce weighted dilated aggregation mechanisms to incorporate the information coming from edge-weights of weighted graphs. Finally, we develop a unified end-to-end encoder-decoder based framework to detect overlapping communities in networks. The deep DynaResGCN model is used as the encoder, whereas we incorporate the Bernoulli-Poisson model as the decoder. Consequently, we apply our overlapping community detection framework in a research topics dataset, a set of networks from Facebook, and in a set of very large co-authorship networks. Our experimentation on these datasets shows significantly superior performance over many state-of-the-art methods for the detection of overlapping communities in networks. Finally, this work has many different important applications, ranging from bioscience to social science and any tasks where there is a network structure.
