Diffuse and Smooth: Beyond Truncated Receptive Field for Scalable and Adaptive Graph Representation Learning