GPU-accelerated Multi-relational Parallel Graph Retrieval for Web-scale Recommendations

TL;DR

This paper addresses the challenge of accurate and scalable retrieval in web-scale recommender systems by proposing GMP-GR, a GPU-accelerated framework that unifies multi-relational user-item relevance learning with a Hierarchically Parallel Graph ANNS. It introduces two key components: MUIRML, a multi-relational metric learning approach with a self-covariant loss to preserve efficient graph navigation, and HiPANNS, a parallel graph search algorithm consisting of inter-candidate and inter-query strategies to handle high concurrency. The authors also integrate system optimizations such as traffic dispersal, compilation optimization, model quantization, and timely data updates, and validate the approach through offline benchmarks and real online deployments at Baidu, achieving high throughput and improved retrieval quality. The work demonstrates that combining multi-relational reasoning with hierarchical graph search and system-level optimizations yields substantial gains in both accuracy and efficiency, with tangible impact in serving hundreds of millions of users at web scale.

Abstract

Web recommendations provide personalized items from massive catalogs for users, which rely heavily on retrieval stages to trade off the effectiveness and efficiency of selecting a small relevant set from billion-scale candidates in online digital platforms. As one of the largest Chinese search engine and news feed providers, Baidu resorts to Deep Neural Network (DNN) and graph-based Approximate Nearest Neighbor Search (ANNS) algorithms for accurate relevance estimation and efficient search for relevant items. However, current retrieval at Baidu fails in comprehensive user-item relational understanding due to dissected interaction modeling, and performs inefficiently in large-scale graph-based ANNS because of suboptimal traversal navigation and the GPU computational bottleneck under high concurrency. To this end, we propose a GPU-accelerated Multi-relational Parallel Graph Retrieval (GMP-GR) framework to achieve effective yet efficient retrieval in web-scale recommendations. First, we propose a multi-relational user-item relevance metric learning method that unifies diverse user behaviors through multi-objective optimization and employs a self-covariant loss to enhance pathfinding performance. Second, we develop a hierarchical parallel graph-based ANNS to boost graph retrieval throughput, which conducts breadth-depth-balanced searches on a large-scale item graph and cost-effectively handles irregular neural computation via adaptive aggregation on GPUs. In addition, we integrate system optimization strategies in the deployment of GMP-GR in Baidu. Extensive experiments demonstrate the superiority of GMP-GR in retrieval accuracy and efficiency. Deployed across more than twenty applications at Baidu, GMP-GR serves hundreds of millions of users with a throughput exceeding one hundred million requests per second.

Figures (9)

• Figure 1: Applications of recommender systems at Baidu.
• Figure 2: An overview of the GPU-accelerated Multi-relational Parallel Graph Retrieval framework.
• Figure 3: Analysis of correlations between retrieval performance and efficiency on Video and Explore datasets. Methods close to the left top ($\nwarrow$) have better retrieval performance, i.e., Cov ($\uparrow$), with less time costs ($\leftarrow$).
• Figure 4: Ablation study of GMP-GR on retrieval performance (Cov) and speed.
• Figure 5: Temporal statistics of traffic and latency.