Abstract: On January 6, 2018, Zhou Jun (alias Xiting) of Ant Financial made a keynote speech entitled “Ant Financial Intelligent Financial Practice” at the Cloud Community Data Intelligent Technology Forum. Is applied more and more of the financial scene, also presents more challenges to financial services, west pavilion introduces the ant gold suit to cope with the challenge of systemic risk prediction and monitoring, based on user semantic actions and recent intelligent assistant and precise marketing and recommend services such as technology, in addition, he also introduced the fee based on AI products.

Here are the highlights:

At present, more and more intelligent technology scenarios are applied, such as micro-loan, insurance, payment, risk control, wealth and so on, which also pose more challenges to financial services, such as time sensitivity, massive data, business diversity, system risk, strong security, automation and so on. In the fields of image/speech, NLP, machine learning, inference and decision, the application of reinforcement learning, unsupervised learning, graph reasoning, transfer learning and other techniques is expected to achieve fast processing and real-time confrontation in the case of large-scale data.

Deep learning + graph: Systematic risk prediction and monitoring

The security of user funds needs to be guaranteed in user accounts, equipment and merchants. The traditional risk control technology is based on rules and policies. With the increase of cases and the increase of rules, the traditional model is difficult to meet the current needs. Ant Financial uses tree model to further judge whether the account is stolen for untrusted transactions. At the same time, GBDT+DNN was adopted to further improve the stolen account model, which has improved the detection rate by 10%. Alipay, for example, allows more than 10 million transactions a day to pass risk checks more quickly and accurately. This is very beneficial to the system itself, the company’s cost and the user’s sense of security. Here is another application of the graph learning model: spam account recognition






Intelligent assistant: intelligent customer service beyond people’s satisfaction


Massive feature extraction based on hash: fast and efficient

A large scale machine learning framework based on parametric server is constructed through deep learning and online learning. The framework has the characteristics of data and model parallelism, robust failover, synchronous and asynchronous iteration, support 100 billion features, 100 billion samples, 1 trillion parameters and so on. In the safe and trusted transaction identification model, the case recall rate increases from 91% to 98% under the same coverage. It allows more than 10 million transactions a day to pass risk checks faster and more accurately. A large-scale matrix factorization algorithm is applied ina typical recommendation scenario, which uses Binary Hash rather than real vector preference, in exchange for significant prediction time and storage savings with a negligible loss of accuracy: 100 million x 10 million matrix factorization converges in 2 hours. The CTR of the word-of-mouth guess you like scenario has a significant increase: CTR of the headline version: 2.5%->5.5, an increase of more than 120%.

Deep reinforcement learning: Temporal decision making — Marketing and recommendation


One-click deployment and effectiveness monitoring





  • Visual modeling interaction: drag-and-drop process composition, auxiliary feature design, complete analysis and evaluation.
  • Rich algorithm access: machine learning deep learning conventional algorithm; Image finance and other vertical domain algorithm; TensorFlow and other open source support.
  • Massive data storage computing: super-large scale parameter server; With GPU,FPGA high performance computing.
  • Global model asset Management: community team collaboration; Feature sharing, experiment sharing; Domain knowledge precipitation, expert experience accumulation.
  • Convenient model deployment: One-click from model completion to release; Strong consistency in offline features is guaranteed.
  • Efficient online forecasting: 99.99% availability guarantee; High-performance floating point and matrix operations; Standardized interface pluggable operator.
  • Exploratory model evolution: support for A/B testing frameworks; Whole life cycle model effect monitoring; Actively discover optimal models and parameters.

Image loss product “Loss treasure”


This article is compiled and edited by Wang Chaoyang, a volunteer group of cloud habitat community, and reviewed by Cheng Tao.

Activity related articles



Ali Cloud Zhu Jintong: Deep secrets of ET brain



Alibaba a wai: customer service link intelligent solutions



Alibaba Shaojie: Big data processing practice



Yang Yang, Zhejiang University: Group User Behavior analysis and presentation learning in social Networks



Alibaba iDST Janssen: The application of intelligent decision in e-commerce platform



Alibaba Bumin: Graph computing is productivity