Application of Graph Neural Networks in Recommendation Systems: OPPO Business Scenario Practice

Graph neural networks (GNNs) have become a core technique for learning graph representations, evolving from early random-walk based methods (DeepWalk, Node2Vec) to spectral approaches (ChebyNets, GCN) and finally to spatial neighbor‑aggregation models that reduce computational complexity.

In recommendation systems, graphs naturally capture user behavior, multi‑type interactions, and cross‑scenario relationships, addressing issues such as data sparsity, cold‑start, and semantic gaps between queries and items. Graph‑based recall can be implemented as an independent route or fused with traditional twin‑tower models, either as additional features or as end‑to‑end sub‑networks.

OPPO’s practical deployment focuses on the app store recommendation scenario. The architecture consists of a classic twin‑tower backbone (user and query towers) augmented on the item side with a graph sub‑network that aggregates one‑hop, two‑hop, and three‑hop neighbor information using an NIA‑GCN‑style aggregation. Separate embeddings are kept for graph edges to preserve distinct relational signals, while query‑related features are shared to align user and item representations.

Experimental results show that adding the graph sub‑network improves both click‑through‑rate (CTR) and ranking metrics (AUC, GAUC), especially boosting the relevance of recommended apps. The approach also enables automatic feature crossing via graph edges, with sparsity regularization (L0) to filter noisy interactions.

Future work aims to build unified pre‑trained graph embeddings across multiple OPPO scenarios (search, advertising, supply‑chain) and to develop noise‑filtering mechanisms that retain useful graph structure while discarding irrelevant connections.