A Regularized Graph Neural Network Based on Approximate Fractional Order Gradients

Graph representation learning is a significant challenge in graph signal processing (GSP). The flourishing development of neural networks (GNNs) provides effective representations for GSP. To effectively learn from signals, we propose regularized network based on approximate fractional order gradients (FGNN). propagates the information between neighboring nodes. approximation strategy calculating derivatives avoids falling into extrema and overcomes high computational complexity derivatives. We further prove that such an feasible FGNN unbiased towards global optimization solution. Extensive experiments citation community show proposed has improved recognition accuracy convergence speed than vanilla FGNN. five datasets different sizes domains confirm great scalability our method.