- Abstract: Graph representation learning serves as the core of many important prediction tasks, ranging from product recommendation in online marketing to fraud detection in financial domain. Real-life graphs are usually large with complex local neighborhood, where each node is described by a rich set of features and easily connects to dozens or even hundreds of neighbors. Most existing graph learning techniques rely on neighborhood aggregation, however, the complexity on real-life graphs is usually high, posing non-trivial overfitting risk during model training. In this paper, we present Neural Sparsification (NeuralSparse), a supervised graph sparsification technique that mitigates the overfitting risk by reducing the complexity of input graphs. Our method takes both structural and non-structural information as input, utilizes deep neural networks to parameterize the sparsification process, and optimizes the parameters by feedback signals from downstream tasks. Under the NeuralSparse framework, supervised graph sparsification could seamlessly connect with existing graph neural networks for more robust performance on testing data. Experimental results on both benchmark and private datasets show that NeuralSparse can effectively improve testing accuracy and bring up to 7.4% improvement when working with existing graph neural networks on node classification tasks.