Industrial-Scale Graph Learning: Boosting Ad ROI and Winning Beijing’s Science Award

On November 7, at the Beijing Science and Technology Awards Conference, the project Industrial‑Scale Graph Learning System Development and Application jointly submitted by Peking University and Alibaba Mama won the second prize of the 2024 Beijing Science and Technology Progress Award. The project was completed by the two research teams under the support of the “Peking University‑Alibaba Mama AI Innovation Joint Laboratory”.

The project results have published over 50 papers in major international journals and conferences, winning awards such as the Best Paper Award at the flagship CIKM conference in knowledge management. The industrial graph learning system has been promoted to millions of merchants across nearly 200 industries nationwide, and over the past three years has been applied to Alibaba Mama’s search advertising system, increasing ad spend scale by more than 3%.

Innovatively, the project proposed a decoupled dynamic graph embedding method theoretically equivalent to the classic Skip‑gram model, reducing algorithm complexity to near‑linear while maintaining embedding accuracy, solving large‑scale dynamic graph embedding challenges; it also introduced a subgraph‑structure‑fusing graph neural network based on anonymous random walks, adding adaptive learning mechanisms for receptive fields and heterogeneous data distributions, and theoretically proved that the model’s expressive power surpasses the limits of graph isomorphism testing, opening new paths for enhancing GNN expressiveness. The team led open‑source work on industrial graph learning frameworks, developing the Euler deep graph learning platform that supports curvature‑space distributed deep modeling and representation learning for large‑scale, complex heterogeneous graphs.

The joint lab, established in September 2022, is a university‑enterprise R&D platform co‑built by Peking University and Alibaba Mama, focusing on AI frontier research such as graph machine learning (graph embedding, GNNs, large‑scale graph pre‑training), decision intelligence (AI economics, game theory, simulation rule making), content creation (intelligent music production), and multimodal intelligence (unified multimodal large‑model fine‑tuning and evaluation).

Since its inception, the lab has produced dozens of top‑conference papers, multiple patents, and successfully applied research to Alibaba Mama’s business, generating significant economic and social value. In addition to the award‑winning graph learning work, the decision intelligence and content creation directions have also achieved notable breakthroughs in both academic research and industrial applications.

For example, the lab’s decision‑intelligence project “Automatic Bidding in Large‑Scale Auctions” was selected for the NeurIPS 2024 conference, and the same‑named competition was successfully held. AIGB (AI‑Generated Bidding) was first proposed by Alibaba Mama in 2023 as a new paradigm for bidding model training. AIGB attracted wide industry attention, and at NeurIPS 2024 a workshop was unusually hosted by a Chinese company.

At NeurIPS 2024, Alibaba Mama officially open‑sourced the AIGB Benchmark and launched a large‑scale auction automatic bidding competition with a dedicated AIGB track, becoming the only domestic industry organization to obtain NeurIPS competition hosting rights last year. Alibaba Mama upgraded AIGB to version AIGB‑R1, further improving ad ROI by 12%; its ad large model LMA2 increased parameter scale tenfold to the trillion‑parameter level.