An integrated fuzzy neural supervision and attention-based graph neural network for improving network clustering

In recent years, graph neural network (GNN) and auto-encoding (AE) have been widely utilized in multiple data mining problems. These architectures have demonstrated powers in data high-dimensional embedding for improving the performance of various task-driven learning tasks, like as clustering. However, most of recent GNN-based cluster techniques still suffered several limitations. These limitations are related to the capability of simultaneously preserving the low-levelled latent feature and global structural representations of the given network/graph. These view-varied graph representations can help to improve the performance of multi-scaled clustering task. Moreover, the achieved multi-viewed structural node embeddings which are learnt by GNN-based architectures might also involve with problems. These problems are related to feature noise and data uncertainty. These feature noise/uncertainties are occurred within representation learning process. These limitations can directly lead to the downgrade in the accuracy performance for clustering tasks. To overcome existing limitations, within this paper, we proposed a novel fuzzy-driven noise-reduced attention-based graph auto-encoding mechanism for network clustering, called as: FAGC. In general, our proposed FAGC model is as an attention-driven multi-layered graph-based AE architecture which is integrated with a custom de-noising fuzzy neural network. In our proposed FAGC model, we integrate the fuzzy neural network with GNN to eliminate problems which are related to the feature uncertainty and noise occurrence during the graph representation learning process. Later, the better quality as well as rich-structural network representations which are generated from our FAGC model are utilized to achieve state-of-the-art performances for the network clustering problem. The extensive experiments within benchmark networked datasets and comparative studies demonstrated the effectiveness and outperformance of our proposed FAGC model in comparing with state-of-the-art graph embedding techniques.