Ant Group’s AI Research Achievements in 2019: Papers, Algorithms, and Applications

2019 was a rapid development year for artificial intelligence and machine learning; Ant Group presented many AI technologies, products, solutions, and research results at top conferences such as NeurIPS, KDD, ICML, SIGMOD, and SIGIR.

Ant Group serves customers far exceeding major banks, with an "AI runs the show" model enabling a wide range of financial services—including consumer credit, wealth management, insurance, and even Ant Forest—without human intervention.

The AI team published papers covering computer vision, natural language processing, machine learning, large‑scale distributed learning, reinforcement learning, graph machine learning, unsupervised learning, and data mining, addressing challenges like dynamic networks, robustness, security, risk, and real‑time processing in financial scenarios.

Notable contributions include a generative adversarial user model for few‑shot reinforcement learning applied to recommendation systems, a particle‑flow Bayes’ rule algorithm for high‑dimensional Bayesian inference, and a new CVRP algorithm that outperforms Google OR‑tools by 10% while leveraging both machine learning and traditional optimization.

To improve model safety and robustness, the team introduced adversarial instance detection (AAAI 2019), identified three attack methods on graph neural networks (ICML 2018), and proposed targeted adversarial training techniques.

The SAFE (Scalable Automatic Feature Engineering) framework (ICDE 2020) was developed to provide high‑throughput, distributed, real‑time feature generation for industrial financial tasks.

In game‑theoretic research, Ant Group proposed Double Neural Counterfactual Regret Minimization (ICLR 2020) and faster CFR variants that dramatically reduce memory usage and accelerate convergence in large‑scale imperfect‑information games.

A value‑propagation algorithm for multi‑agent reinforcement learning was created to solve decentralized training and execution challenges, with applications in fund optimization and traffic control.

Further advances include a DROP‑ranking system that uses graph construction and Hard‑EM for numeric reasoning in NLP, and a reinforcement‑learning‑based intent prediction method showcased at SIGIR 2019 for “answer before asked” capabilities.

All these intelligent services are underpinned by OceanBase, Ant Group’s financial‑grade distributed relational database, which topped the TPC‑C benchmark in 2019 and now supports both internal services and dozens of external financial institutions.

The combination of cutting‑edge research and real‑world deployment demonstrates how AI drives inclusive finance, from voice authentication and smart waste classification to graph‑based marketing and risk control, while promotional material invites readers to download technical e‑books and follow the Ant Group technology public account.