An integrated simplicial neural network with neuro-fuzzy network for graph embedding

In recent years, graph neural network (GNN) has become the main stream for most of recent researches due to its powers in dealing with complex graph data learning problems. However, as most of the recent GNN-based architectures have been mainly designed to only evaluate direct relational structures between nodes. As the results, these techniques are unable to capture the sophisticated multi-way relationships in graph. The multi-way relationships can be represented by both explicit graph-based and complex topological structures. In general, the multi-way relationships in graph can be modelled as simplicial complexes, hyper-graphs, e.g., and can be efficiently preserved under the simplicial neural networks (SNN). There are several notable SNN-based architectures have been proposed recently, such as the well-known simplicial convolutional neural network (SCNN). The SNN-based techniques have shown the competitive performances in handling graph learning. However, most of recent proposed SNN-based architectures are designed upon the deep neural learning paradigm. Therefore, they still encountered several challenges with regard to the feature noise and data uncertainty. To overcome these limitations, in this paper, we proposed a novel integrated SNN and neuro-fuzzy network (NFN) graph embedding technique, called as: SFGE. Our SFGE model is designed to better capture the multi-way structural representation of graph by taking advances of different advanced graph-based and fuzzy-based neural learning techniques. By taking advances of neuro-fuzzy learning approach, our model can efficiently support to eliminate the feature uncertainty/ambiguity during the task-driven fine-tuning process. In addition, it also supports to better capture the rich multi-way relational structures of the input graphs under the topology-enhanced graph analysis approach. Extensive empirical studies within a real-world molecular graph dataset have effectiveness of our SFGE in dealing with graph classification task.