Ant Group Unveils Large Graph Model (LGM) Merging Graph Computing with Large Language Models

Bill Gates recently described generative AI as the most important technological revolution in the past 40 years, highlighting the industry’s focus on how such intelligence can drive scientific research and innovative applications.

On September 7, 2023, at the 2023 Bund Conference sub‑forum “New Generation Data Infrastructure – Exploring Graph Intelligence Applications and Development,” Ant Group introduced the Large Graph Model (LGM). This research integrates graph computing, graph learning, and large language models, leveraging the generative power of LLMs and the relational analysis capability of graph computation to deliver more intuitive, comprehensive information and precise insights, thereby addressing the challenges of massive, complex digital applications. Ant has completed the first phase, termed “generative heterogeneous graph augmentation,” and the related paper was accepted at the WWW 2023 conference.

Graph computing is a powerful data‑processing technique that resolves relational challenges in complex networks and finds applications in financial fraud detection, weather forecasting, drug discovery, and even brain‑inspired research, earning it the nickname “the nose of AI.” Large models are viewed as the most promising path toward artificial general intelligence, achieving human‑level or superior performance in certain tasks.

Why use cutting‑edge technology to drive cutting‑edge technology? According to Ant senior technical expert Liu Yongchao, large language models can infer hidden relationships but cannot render relationship graphs, which are essential for clear data‑relationship visualization. Combining LLMs with graph computing first performs logical reasoning on massive information and then uses super‑computing to calculate relationships, akin to attaching a supercomputer to the human brain.

The Ant Group’s research consists of two main components. First, it employs large language models to enrich graph data—a process called “generative heterogeneous graph augmentation,” where LLMs generate new data points to expand and diversify multi‑type graph datasets. Second, it uses prompts (specific instructions) to guide the model in learning and discovering particular data characteristics; for example, a prompt like “common features of groups with more than three defaults in a year” enables the model to generate samples that satisfy the condition, accelerating data analysis and feature discovery.

Ant Group is a leader in graph computing, co‑developing the TuGraph platform with Tsinghua University, breaking the LDBC SNB world record three times, winning the “Leading Technology Achievement” award at the 2021 World Internet Conference, and being named a leader in the IDC MarketScape China Graph Database market in 2023. The LGM project began in early 2022 and, following the surge of large models at the end of 2022, has demonstrated feasibility.

During the forum, experts and scholars shared additional advances: Professor M. Tamer Özsu from the University of Waterloo discussed streaming graph computing; Professor Chen Huajun from Zhejiang University highlighted knowledge processing opportunities and challenges in the era of large models; Deputy Director Chen Hongyang from Zhijiang Laboratory presented the latest research on graph scientific computing for biomedicine; and Li Yashou, co‑founder and deputy editor of Machine Heart, expressed optimism that combining graph intelligence with large models will significantly boost data intelligence.

The discussion underscored the promising synergy between artificial intelligence and graph computing, outlining a vital development trajectory for graph intelligence.