This paper introduces Simple Graph Convolution (SGC), a simplified version of Graph Convolutional Networks (GCNs). SGC removes nonlinearities and collapses weight matrices between layers, resulting in a linear model that corresponds to a fixed low-pass filter followed by a linear classifier. Theoretical analysis shows that SGC is equivalent to a fixed filter in the graph spectral domain, which produces smooth features across the graph. Experimental results demonstrate that SGC achieves comparable or superior performance to GCNs on various tasks while being significantly faster and more efficient. SGC is computationally more efficient, scales to larger datasets, and is naturally interpretable. It outperforms FastGCN by up to two orders of magnitude on the largest dataset, Reddit. SGC is also effective on a wide range of downstream tasks, including text classification, user geolocation, relation extraction, and zero-shot image classification. The model is simple, efficient, and interpretable, making it a valuable baseline for future graph learning research. SGC's performance highlights the effectiveness of graph convolution filters and suggests that the expressive power of GCNs may primarily come from repeated graph propagation rather than nonlinear feature extraction.This paper introduces Simple Graph Convolution (SGC), a simplified version of Graph Convolutional Networks (GCNs). SGC removes nonlinearities and collapses weight matrices between layers, resulting in a linear model that corresponds to a fixed low-pass filter followed by a linear classifier. Theoretical analysis shows that SGC is equivalent to a fixed filter in the graph spectral domain, which produces smooth features across the graph. Experimental results demonstrate that SGC achieves comparable or superior performance to GCNs on various tasks while being significantly faster and more efficient. SGC is computationally more efficient, scales to larger datasets, and is naturally interpretable. It outperforms FastGCN by up to two orders of magnitude on the largest dataset, Reddit. SGC is also effective on a wide range of downstream tasks, including text classification, user geolocation, relation extraction, and zero-shot image classification. The model is simple, efficient, and interpretable, making it a valuable baseline for future graph learning research. SGC's performance highlights the effectiveness of graph convolution filters and suggests that the expressive power of GCNs may primarily come from repeated graph propagation rather than nonlinear feature extraction.