Baidu's PGL2.2: A Graph Neural Network Framework, Techniques, and Real‑World Applications

Graphs provide a universal language for representing complex relationships in social networks, proteins, recommendation systems, and more. Graph neural networks (GNNs) extend neural networks to graph structures using a message‑passing paradigm.

Baidu developed PGL2.2 on top of the Paddle deep‑learning framework, exposing APIs for graph construction, sampling, tensorization, and various graph models (walk‑based, message‑passing, knowledge‑embedding). The system supports both static and dynamic graph modes, enabling users to define node and edge tensors and customize message flow.

Training strategies are categorized by graph size: small graphs fit in a single GPU and use full‑batch training; medium graphs exceed a single GPU’s memory and employ graph‑partitioning or multi‑GPU chunked training with NCCL synchronization; large graphs adopt mini‑batch sampling with neighbor sampling and distributed execution, often requiring parameter‑server‑based embedding storage.

Key GNN tasks covered include node classification (semi‑supervised, using label propagation and the UniMP algorithm for label‑feature fusion) and link prediction (addressing depth and neighbor‑sampling bottlenecks with relation‑aware embedding propagation and the REP technique). The presented methods achieved state‑of‑the‑art results on Open Graph Benchmark datasets and won the KDDCup 2021 Wiki90M competition.

Industrial applications of PGL span Baidu Search (web‑page quality assessment, anti‑spam), recommendation systems, risk control, traffic prediction in Baidu Maps, and POI retrieval. The framework abstracts graph types (homogeneous, heterogeneous, bipartite), sampling strategies (node2vec, meta‑path), and GNN aggregators, supporting large‑scale sparse features and automatic heterogeneous‑graph extensions such as LightGCN‑relation‑wise.

A short Q&A highlights communication overhead in distributed training, the use of multi‑head attention to mitigate over‑smoothing, and successful integration of GNNs with language models (e.g., ERINESage) for NLP‑related tasks.