AdaGNN

Abstract

Graph Neural Networks have recently become a prevailing paradigm for various high-impact graph analytical problems. Existing efforts can be mainly categorized as spectral-based and spatialbased methods. The major challenge for the former is to find an appropriate graph filter to distill discriminative information from input signals for learning. Recently, myriads of explorations are made to achieve better graph filters, e.g., Graph Convolutional Network (GCN), which leverages Chebyshev polynomial truncation to seek an approximation of graph filters and bridge these two families of methods. Nevertheless, it has been shown in recent studies that GCN and its variants are essentially employing fixed low-pass filters to perform information denoising. Thus their learning capability is rather limited and may over-smooth node representations at deeper layers. To tackle these problems, we develop a novel graph neural network framework AdaGNN with a well-designed adaptive frequency response filter. At its core, AdaGNN leverages a simple but elegant trainable filter that spans across multiple layers to capture the varying importance of different frequency components for node representation learning. The inherent differences among different feature channels are also well captured by the filter. As such, it empowers AdaGNN with stronger expressiveness and naturally alleviates the over-smoothing problem. We empirically validate the effectiveness of the proposed framework on various benchmark datasets. Theoretical analysis is also provided to show the superiority of the proposed AdaGNN. The open-source implementation of AdaGNN can be found here: https://github.com/yushundong/AdaGNN.

Graph Neural Networks with Diverse Spectral Filtering

“…To alleviate this issue, we introduce a diverse spectral filtering (DSF) framework to enhance spectral GNNs with trainable diverse filters. It is worth noting that though both JacobiConv [42] and AdaGNN [12] learn multiple filters in a seemingly similar way, they are essentially different from our method in nature. Concretely, their adaptive filters are mainly for studying each feature channel independently, whilist our diverse filters aim at individual context modeling for each node.…”

“…Recent studies have shown that most spectral GNNs operate as polynomial spectral filters [9,20,42] with either fixed designs such as GCN [24], APPNP [15], and GNN-LF/HF [51] or learnable forms, e.g., ChebNet [11], AdaGNN [12], GPR-GNN [9], ARMA [5], Bern-Net [20], and JacobiConv [42]. Further remarks can be found in Appendix A.…”

“…In the second class, spectral GNNs are mostly designed with trainable filters: ChebNet [11] approximates the filtering operation with Chebyshev polynomials whose coefficients are learnable. AdaGNN [12] learns adaptive filters to model each feature channel independently. GPR-GNN [9] extends APPNP [15] by directly parameterizing its filter weights and training them with gradient descent.…”

SocialCU: Integrating Commonalities and Uniqueness of Users and Items for Social Recommendation

“…The major applications of GNN in SR are models based on Graph Convolutional Networks (GCN) [30][31][32] and Graph Attention Networks (GAT) [33]. These models use the information propagation mechanism to capture higher-order relationships between nodes and obtain the influence of different neighboring nodes to form the final representation for recommendation tasks [34,35]. Depending on how the data is modeled, GNN applications can be assigned to two categories.…”

CurvAGN: Curvature-based Adaptive Graph Neural Networks for Predicting Protein-Ligand Binding Affinity

“…To address this issue, several GNN designs have been proposed that are specifically tailored for heterophilic graphs. These designs include MixHop [50], MM-DAN [51], BeyondGNN [32], AdaGNN [52], Beyond-GCN [53], and Geom-GCN [54]. Persistent curvature descriptors have been shown to be effective in representing protein-ligand complexes, but they rely on prior knowledge.…”