Financial Graph Machine Learning, AutoML, and Multi‑Agent Reinforcement Learning at Ant Financial

On September 27, at the Cloudwise Conference in Hangzhou, AI expert Professor Song Le, a researcher at Ant Financial and associate professor at Georgia Tech, discussed the development and application of financial‑specific machine learning within Ant Financial.

Deep Learning System for Massive Graph Data – Financial data forms a huge heterogeneous graph of nodes (users, merchants, accounts, devices, locations) and edges representing fund flows. Modeling such graphs requires converting irregular graph structures into vector representations via a dedicated platform, enabling downstream models such as recommendation, decision‑making, and generative models.

Ant Financial’s graph data can reach hundreds of billions of nodes and trillions of edges, demanding robust storage, fast query, sub‑graph extraction, and distributed computation. The platform provides generic operators for sampling, random walks, and message passing, supporting both unsupervised and supervised representation learning, which in turn powers predictions of node/edge types, temporal behaviors, and multi‑objective outcomes for various financial services.

Algorithm Library for Multi‑Type Graphs

xGrep – scalable training on attribute‑free graphs with billions of nodes and edges, featuring distributed random‑walk and word2vec frameworks.

GeniePath – adaptive graph neural network for attribute graphs, delivering industry‑leading performance.

HeGNN & IGNN – hierarchical attention mechanisms for heterogeneous graphs, offering financial‑grade interpretability.

KGNN – knowledge‑graph embedding combining GNNs with graph‑based models.

Interpretability is crucial in finance; models such as GeniePath can trace which neighboring nodes or edges influence a risk‑control decision, while HeGNN/IGNN expose multi‑level and multi‑dimensional impacts.

Automated Machine Learning (AutoML) System – To alleviate the manual effort of model design and hyper‑parameter tuning, Ant Financial built an AutoML platform that automates feature engineering, hyper‑parameter search, neural architecture search, and meta‑learning, enabling rapid model iteration across thousands of daily training jobs.

A concrete case, “autonet,” automatically assembles proven DNN sub‑modules into new architectures, achieving comparable training time to manually designed models while improving conversion rates by 14% in user acquisition scenarios.

Multi‑Agent Reinforcement Learning System – Financial nodes are dynamic and engage in competitive or cooperative interactions. Traditional RL assumes a static simulator, which is unsuitable for finance. Ant Financial therefore constructed a multi‑agent RL platform that learns user behavior simulators and reward functions via imitation learning, enabling large‑scale RL for tasks such as fraud detection and personalized recommendation.

One application integrates RL into a recommendation system, treating user interactions as a long‑term optimization problem rather than isolated predictions, thereby increasing sustained user engagement.

These technologies have been demonstrated in real‑world deployments: graph deep learning improves marketing cost efficiency by 8% and adds hundreds of billions of credit limits; AutoML accelerates model development across business units; and multi‑agent RL expands to fraud, payment, and other fintech scenarios.

Overall, Ant Financial’s high‑availability graph learning platform combines system engineering, high‑performance computing, and advanced AI algorithms to support diverse financial applications.