NodeDependentLocalSmoothingforScalable GraphLearning

Neural Information Processing Systems 

To make the proof concise, we will assume matrixP is connected, otherwise we can perform the same operation inside each block. With the help of NDLS, Random Forest and XGBoost outperforms their base models by6.1% and 7.5% respectively. In these three networks, papers from different topics are considered asnodes, and the edges are citations among the papers. Industry is a short-form video graph, collected from a real-world mobile application from our industrial cooperativeenterprise. Wesampled 1,000,000 users and videos from the app, and treat these items as nodes.