Graph Attention Multi‑Layer Perceptron (GAMLP) and Node‑Dependent Local Smoothing (NDLS) for Scalable and Flexible Graph Neural Networks

In real‑world scenarios graph data is massive and heterogeneous, leading to scalability and flexibility challenges for traditional GNNs.

The talk introduces two Tencent Angel Graph solutions: Node‑Dependent Local Smoothing (NDLS) and Graph Attention Multi‑Layer Perceptron (GAMLP). NDLS determines an optimal propagation depth for each node by comparing its intermediate features to a steady‑state feature, thereby mitigating over‑ and under‑smoothing.

GAMLP builds on NDLS by performing node‑wise feature and label propagation, then combines them with attention mechanisms (Recursive Attention and JK‑Attention) before feeding the result to an MLP. The model is fully decoupled, highly scalable, and can be used with any downstream predictor.

Extensive experiments on transductive and inductive benchmarks (including OGB datasets) show that both NDLS and GAMLP consistently outperform baselines, achieve deep propagation without over‑smoothing, and maintain competitive training speed even on billion‑node graphs.

The conclusions highlight the methods’ advantages in scalability, flexibility, and efficiency, and discuss remaining limitations such as the lack of end‑to‑end training and the simple averaging of multi‑hop features.