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

This article introduces the background of graph neural networks (GNN) and graph representation learning, describing how vectors are used to represent nodes, edges, and whole graphs.

It reviews traditional graph embedding methods such as random‑walk based models (DeepWalk, Node2Vec) and spectral graph convolution approaches (ChebyNets, GCN), highlighting their advantages and limitations.

The article explains why graphs are valuable for recommendation systems, enumerating four benefits: the natural graph structure of user behavior, the ability to incorporate heterogeneous actions, linking multiple scenarios, and alleviating sparsity and cold‑start problems.

It compares conventional recommendation pipelines (dual‑tower models) with graph‑based approaches for recall and ranking, describing graph‑based recall (independent graph route and graph‑fusion) and graph‑enhanced ranking (graph features and graph sub‑networks).

The OPPO case study details the architecture (data, platform, algorithm, application layers) and the specific challenges of app‑store recommendation—high relevance requirements, long‑tail queries, and semantic gaps—and presents a hybrid model that adds a graph sub‑network to the item side of a dual‑tower architecture, using multi‑hop neighbor aggregation (NIA‑GCN) and adaptive fusion.

Experimental results show that the graph sub‑network improves click‑through rate and ranking metrics (AUC, GAUC), especially for items, and the article concludes with future directions such as unified multi‑scenario pre‑training and noise filtering in large graphs.