DC-GNN: Decoupled Graph Neural Networks for Improving and Accelerating Large-Scale E-commerce Retrieval

Abstract

In large-scale E-commerce retrieval, the Graph Neural Networks (GNNs) has become one of the stage-of-the-arts due to its powerful capability on topological feature extraction and relational reasoning. However, the conventional GNNs-based large-scale E-commerce retrieval suffers from low training efficiency, as such scenario normally has billions of entities and tens of billions of relations. Under the limitation on efficiency, only shallow graph algorithms can be employed, which severely hinders the GNNs representation capability and consequently weakens the retrieval quality. In order to deal with the trade-off between training efficiency and representation capability, we propose the Decoupled Graph Neural Networks (DC-GNN) to improve and accelerate the GNNs-based large-scale E-commerce retrieval. Specifically, DC-GNN decouples the conventional framework into three stages: pre-train, deep aggregation, and CTR prediction. By decoupling the graph operations and the CTR prediction, DC-GNN can effectively improve the training efficiency. More importantly, it can enable deeper graph operations to adequately mine higher-order proximity to boost model performance. Extensive experiments on large-scale industrial datasets demonstrate that DC-GNN gains significant improvements in both model performance and training efficiency.