Artificial Intelligence 13 min read

AI Algorithm Practices and Data Platform Architecture at Ping An Bank

The article presents Ping An Bank's AI-driven data platform, covering business background, architectural layers, algorithmic applications such as customer segmentation, portrait, business forecasting, and graph analysis, and shares practical insights on platform design, model deployment, and the role of data product managers.

Ctrip Technology

Jan 18, 2018

AI Algorithm Practices and Data Platform Architecture at Ping An Bank

Author Introduction

Pan Pengju, head of the AI algorithm and analytics team of Ping An Bank's big data platform, previously worked at Ctrip where he built algorithmic engineering frameworks and led teams to improve hotel services. He joined Ping An in 2017 to promote AI applications in banking.

Background

Traditional banks face intense competition from internet and P2P firms, prompting retail transformation and fintech adoption. Over the past year the bank has experimented with algorithms and reflects on future applications.

Business Background

The bank's core KPIs are AUM (assets under management) and LUM (loans under management). AUM drives asset size, while LUM focuses on credit risk; regulatory limits require LUM/AUM ratios to stay below certain thresholds. Challenges include regulation, compliance, customer acquisition difficulty, and sparse data.

Algorithm Practice

The data application architecture consists of three layers: infrastructure (Hadoop, Hive, Spark, Elasticsearch, Redis), data service layer (common products and services), and application layer (AI platforms). The AI platform is containerized with Docker for one‑click training and deployment.

AI capabilities are divided into deep learning and algorithmic applications. Use cases include knowledge graphs for risk, image/voice/text processing for automation, and custom development where open‑source solutions lag behind market products.

Case 1: Customer Management

Customers are segmented by lifecycle (acquisition, migration, churn) and business dimensions, enabling tailored offers and benefits. Sub‑models generate precise customer profiles stored in a tagging system for easy retrieval. Fine‑grained operations can improve efficiency by up to five times compared to coarse approaches.

Case 2: Customer Portrait

A front‑line interface displays factual tags and algorithmic predictions, including churn probability and key influencing factors, as well as model accuracy comparisons to support decision making.

Case 3: Business Forecast

A three‑day volume forecast informs staffing schedules. Due to limited historical data, a model‑fusion approach combining multiple prediction methods and rule‑based adjustments (e.g., for month‑start anomalies) achieved less than 9% absolute error.

Case 4: Graph Analysis

Graph clustering of card numbers, application IDs, and activation numbers identified potential fraud groups. Knowledge graphs built from these clusters enable relationship queries across phones and cards, aiding risk investigation. Graph databases such as Neo4j, OrientDB, and JanusGraph have been evaluated, with OrientDB currently used for POI storage.

Some Reflections

1) Data product managers are essential to bridge business, data, and algorithm teams, streamlining project execution. 2) A unified AI platform that integrates model training, deployment, and visualization reduces duplicated effort. 3) Future platform designs should support end‑to‑end pipelines from data ingestion to online/offline serving, similar to Uber's model.