Key Insights from Ant Financial VP Dr. Qi Yuan’s Talk on the Development and Application of Financial Intelligence at CCAI 2017

On July 22‑23, 2017, under the guidance of the China Association for Science and Technology and the Chinese Academy of Sciences, the 2017 China Artificial Intelligence Conference (CCAI 2017) was held in Hangzhou International Conference Center, organized by the Chinese Association for Artificial Intelligence, Alibaba Group & Ant Financial, and co‑hosted by CSDN and the Institute of Automation, Chinese Academy of Sciences.

During the conference, Ant Financial Vice President and Chief Data Scientist Dr. Qi Yuan delivered a keynote titled “The Development and Application of Financial Intelligence.” He emphasized Ant Financial’s two keywords for the year—“Open” and “AI”—and illustrated how AI underpins products such as risk‑control systems, intelligent assistants, and the “Damage‑Assessment Treasure” (定损宝).

Basic Challenges of Financial Services

Ant Financial argues that AI is indispensable for fintech. The main challenges include:

Time Sensitivity: Making millisecond‑level risk decisions.

Massive Data: Processing billions of daily transactions.

Business Diversity: Leveraging transfer learning to find commonalities across tasks.

Systemic Risk: Analyzing the financial network from a graph perspective.

Strong Data Security: Protecting user privacy and financial data.

Technical Elements of Financial Intelligence

To address these challenges, Ant Financial built a Financial Intelligence Platform that stacks low‑level image understanding and speech recognition, then adds Natural Language Processing (NLP), followed by machine‑learning and deep‑learning models for time‑series analysis (e.g., predicting Yu‑e‑Bao interest rates). At the top layer, reasoning and decision‑making capabilities help users and partners make informed choices.

The platform incorporates reinforcement learning, unsupervised learning, graph reasoning, and federated learning, satisfying the requirements of real‑time, large‑scale, and secure encryption in finance.

Ant Financial positions itself as a “TechFin” company, emphasizing that every mature technology will be opened to partners and the broader financial ecosystem.

Case 1: Secure Risk Control

Traditional risk control relied on rule‑based models. Ant Financial upgraded this by adopting a GBDT + DNN architecture (inspired by Facebook’s 2014 CTR prediction work) using its own parameter‑server framework. Features generated by GBDT are fed into a deep network, improving detection efficiency. Further, embedding techniques construct a graph of users, merchants, and devices, and a Struc2vec‑based model replaces the older Node2vec system, achieving a significant boost in prediction accuracy.

Case 2: Intelligent Customer Service Assistant

Ant Financial’s smart assistants (e.g., “Xiao Ma Da” in Alipay and “Anna” in Ant Fortune) use LSTM + DSSM to match user queries with historical behavior, achieving a 97% self‑service rate during Double‑11. Sentiment analysis combines a financial‑domain knowledge base with deep learning (CNN + TNN), reaching an 88.4% accuracy on news‑text sentiment.

Case 3: Large‑Scale Machine Learning via Parameter Server

Ant Financial extended Alibaba’s parameter‑server‑based machine‑learning platform to support new algorithms. The system now powers Alibaba’s ad search, mobile‑Taobao recommendation, and real‑time recommendations during Double‑11. In Ant’s risk‑control scenario, the same coverage yields a recall increase from 91% to 98% and processes over ten million transactions per day. The platform’s achievements were presented at KDD 2017 and WWW 2017.

Case 4: Reinforcement Learning for Marketing – Huabei Smart Signing

The marketing problem of selecting coupons, channels, and audiences is tackled with a deep reinforcement‑learning framework. The pipeline defines State (user features), Action (card + channel decision), and Reward (click‑through and signing behavior). The approach increased click‑through rate by 171% and signing rate by 149%.

Case 5: Model Service Platform ("Crystal Ball")

Ant Financial built a visual model‑service platform that allows users to train, deploy, and monitor models with drag‑and‑drop operations. The platform supports A/B testing, full‑life‑cycle monitoring, multi‑team collaboration, and treats both data and models as assets.

Case 6: Damage‑Assessment Treasure (定损宝)

The “Damage‑Assessment Treasure” uses computer‑vision techniques to automatically evaluate vehicle damage from photos. Challenges addressed include component recognition, de‑glare, and angle correction. Deep learning drives object detection, segmentation, multi‑image fusion, and decision making, reducing average processing cost to ¥150 and cutting labor by 50%.

Conclusion

Dr. Qi Yuan reflected on the gap between academia and industry, urging closer collaboration because industry possesses abundant data and real‑world challenges. Ant Financial commits to the principle “open a technology once it matures,” hoping its AI capabilities will be openly shared to benefit partners, customers, and society at large.