Macro Graph of Experts for Billion-Scale Multi-Task Recommendation

A research team has introduced the Macro Graph of Experts (MGOE) framework, a graph neural network architecture designed for billion-scale multi-task recommendation. The system has been deployed within Alibaba’s recommender infrastructure, according to a paper posted on arXiv [1]. Traditional multi-task learning methods for recommendation often rely solely on individual user and item embeddings, neglecting the distinct billion-scale graph structures that correspond to different tasks [1]. The MGOE framework addresses this gap by leveraging macro graph embeddings to capture task-specific macro features while modeling correlations between task-specific experts [2]. The architecture introduces a Macro Graph Bottom, which enables multi-task learning models to incorporate graph information, and a Macro Prediction Tower that dynamically integrates macro knowledge across tasks [3]. A key component, the Macro Task Merging Graph (MTMG), merges multiple billion-scale graphs into a single unified macro structure, consolidating complex graph data without introducing what the authors describe as unbearable computational complexity [4]. The paper states that MGOE is the first graph neural network architecture for billion-scale multi-task recommendation to achieve this [5]. The framework was submitted to arXiv in June 2025 and last revised in June 2026 [1]. Authors include Zijin Hong, along with researchers holding substantial publication records such as Hao Chen, who has an h-index of 83 and over 43,000 citations, and Feiran Huang, with an h-index of 31 and more than 3,000 citations [3]. Offline experiments were conducted on three public benchmark datasets, where MGOE demonstrated superiority over state-of-the-art multi-task learning methods [2]. Online A/B tests further confirmed the framework’s performance in a live billion-scale recommender system [1]. The work was developed under arXivLabs, a framework that allows collaborators to build and share new features on the arXiv platform [1].