LtrGCN: Large-Scale Graph Convolutional Networks-Based Learning to Rank for Web Search

While traditional Learning to Rank (LTR) models use query-webpage pairs to perform regression tasks to predict the ranking scores, they usually fail to capture the structure of interactions between queries and webpages over an extremely large bipartite graph. In recent years, Graph Convolutional Neural Networks (GCNs) have demonstrated their unique advantages in link prediction over bipartite graphs and have been successfully used for user-item recommendations. However, it is still difficult to scale-up GCNs for web search, due to the (1) extreme sparsity of links in query-webpage bipartite graphs caused by the expense of ranking scores annotation and (2) imbalance between queries (billions) and web-pages (trillions) for web-scale search as well as the imbalance in annotations. In this work, we introduce the Q-subgraph and W-subgraph to represent every query and webpage with the structure of interaction preserved, and then propose LtrGCN-an LTR pipeline that samples Q-subgraphs and W-subgraphs from all query-webpage pairs, learns to extract features from Q-subgraphs and W-subgraphs, and predict ranking scores in an end-to-end manner. We carried out extensive experiments to evaluate LtrGCN using two real-world datasets and online experiments based on the A/B test at a large-scale search engine. The offline results show that LtrGCN could achieve Delta NDCG(5) = 2.89%-3.97% compared to baselines. We deploy LtrGCN with realistic traffic at a large-scale search engine, where we can still observe significant improvement. LtrGCN performs consistently in both offline and online experiments.