Broad Graph Attention Network With Multiple Kernel Mechanism

Abstract

Graph neural networks (GNNs) are highly effective models for tasks involving non-Euclidean data. To improve their performance, researchers have explored strategies to increase the depth of GNN structures, as in the case of convolutional neural network (CNN)-based deep networks. However, GNNs relying on information aggregation mechanisms typically face limitations in achieving superior representation performance because of deep feature oversmoothing. Inspired by the broad learning system, in this study, we attempt to avoid the feature oversmoothing issue by expanding the width of GNNs. We propose a broad graph attention network framework with a multikernel mechanism (BGAT-MK). In particular, we propose the construction of a broad GNN using multikernel mapping to generate several reproducing kernel Hilbert spaces (RKHSs), where nodes can wander through different kernel spaces and generate representations. Furthermore, we construct a broader network by aggregating representations in different RKHSs and fusing adaptive weights to aggregate the original and enhanced mapped representations. The efficacy of BGAT-MK is validated through experiments on conventional node classification and light detection and ranging point cloud semantic segmentation tasks, demonstrating its superior performance.