Adaptive Kernel Graph Neural Network

Graph neural networks (GNNs) have demonstrated great success in representation learning for graph-structured data. The layer-wise graph convolution GNNs is shown to be powerful at capturing topology. During this process, are usually guided by pre-defined kernels such as Laplacian matrix, adjacency or their variants. However, the adoptions of may restrain generalities different graphs: mismatch between and kernel would entail sub-optimal performance. For example, that focus on low-frequency information not achieve satisfactory performance when high-frequency significant graphs, vice versa. To solve problem, paper, we propose a novel framework - i.e., namely Adaptive Kernel Neural Network (AKGNN) which learns adapt optimal unified manner first attempt. In proposed AKGNN, design data-driven mechanism, adaptively modulates balance all-pass low-pass filters modifying maximal eigenvalue Laplacian. Through AKGNN threshold high low frequency signals relieve generality problem. Later, further reduce number parameters parameterization trick enhance expressive power global readout function. Extensive experiments conducted acknowledged benchmark datasets promising results demonstrate outstanding our comparison with state-of-the-art GNNs. source code publicly available at: https://github.com/jumxglhf/AKGNN.