ABSTRACT
Graph neural networks (GNNs) are widely used for analyzing graph-structural data and solving graph-related tasks due to their powerful expressiveness. However, existing off-the-shelf GNN-based models usually consist of no more than three layers. Deeper GNNs usually suffer from severe performance degradation due to several issues including the infamous "over-smoothing" issue, which restricts the further development of GNNs. In this article, we investigate the over-smoothing issue in deep GNNs. We discover that over-smoothing not only results in indistinguishable embeddings of graph nodes, but also alters and even corrupts their semantic structures, dubbed semantic over-smoothing. Existing techniques, e.g., graph normalization, aim at handling the former concern, but neglect the importance of preserving the semantic structures in the spatial domain, which hinders the further improvement of model performance. To alleviate the concern, we propose a cluster-keeping sparse aggregation strategy to preserve the semantic structure of embeddings in deep GNNs (especially for spatial GNNs). Particularly, our strategy heuristically redistributes the extent of aggregations for all the nodes from layers, instead of aggregating them equally, so that it enables aggregate concise yet meaningful information for deep layers. Without any bells and whistles, it can be easily implemented as a plug-and-play structure of GNNs via weighted residual connections. Last, we analyze the over-smoothing issue on the GNNs with weighted residual structures and conduct experiments to demonstrate the performance comparable to the state-of-the-arts.
ABSTRACT
Graph Neural Networks (GNNs) have demonstrated great potential in achieving outstanding performance in various graph-related tasks, e.g., graph classification and link prediction. However, most of them suffer from the following issue: shallow networks capture very limited knowledge. Prior works design deep GNNs with more layers to solve the issue, which however introduces a new challenge, i.e., the infamous over-smoothness. Graph representation over emphasizes node features but only considers the static graph structure with a uniform weight are the key reasons for the over-smoothness issue. To alleviate the issue, this paper proposes a Dynamic Weighting Strategy (DWS) for addressing over-smoothness. We first employ Fuzzy C-Means (FCM) to cluster all nodes into several groups and get each node's fuzzy assignment, based on which a novel metric function is devised for dynamically adjusting the aggregation weights. This dynamic weighting strategy not only enables the intra-cluster interactions, but also inter-cluster aggregations, which well addresses undifferentiated aggregation caused by uniform weights. Based on DWS, we further design a Structure Augmentation (SA) step for addressing the issue of underutilizing the graph structure, where some potentially meaningful connections (i.e., edges) are added to the original graph structure via a parallelable KNN algorithm. In general, the optimized Dynamic Weighting Strategy with Structure Augmentation (DWSSA) alleviates over-smoothness by reducing noisy aggregations and utilizing topological knowledge. Extensive experiments on eleven homophilous or heterophilous graph benchmarks demonstrate the effectiveness of our proposed method DWSSA in alleviating over-smoothness and enhancing deep GNNs performance.
