Neighborhood-Aware Fuzzy Graph Convolutional Networks for Improved Information Fusion

Graph Convolutional Networks (GCNs) have shown significant performance in various domains. However, their performance is limited by the reliance on static adjacency matrices, which fail to capture the nuanced relationships between nodes, especially when these matrices are limited to coarse binary edges, thereby overlooking the inherent uncertainties and fuzzy connections within the data. Moreover, many existing algorithms for information systems overly rely on explicit features of nodes, and fail to make full use of and mine the topology structure to learn topological features, leading to the omission of critical information. In this paper, we propose the Neighborhood-Aware Fuzzy Graph Convolutional Network (NAFGCN) that dynamically adjusts the adjacency matrix based on neighborhood-aware thresholds. Initially, based on the concept of three-way decisions, thresholds are established in accordance with the context of each node's neighborhood. Subsequently, these thresholds are employed to enhance or attenuate the connections between nodes, thereby constructing a neighborhood-aware fuzzy adjacency matrix. Secondly, we have designed a layered graph convolution method to learn the topological features in information systems, which are then integrated with the node features. Finally, the efficacy of the algorithm is substantiated through comparative and ablation experiments, and its non-random nature is affirmed by the parameters.