Personalized Recommendation System of 51 Credit Card: Architecture, Challenges, and Growth Cases

Artificial intelligence is a key driver of financial industry transformation; this article introduces how 51 Credit Card builds a model system, extracts data value, and creates a private scanning platform for personalized recommendations.

Initially focused on bill management, 51 Credit Card Manager now spans debt management, finance, and technology services.

Operational growth pain points: unlike e‑commerce where recommendation emphasizes recall, finance emphasizes ranking; high‑value customers must be offered alternative products when their primary goal (e.g., applying for a card) is not met.

Key operational challenges: (1) Identifying differentiated user needs as business expands; (2) High acquisition cost of new users requiring precise distribution; (3) Low conversion rates of push channels needing improvement.

Technical difficulties: (1) Sparse financial user behavior makes tag extraction critical; (2) Building a real‑time preference prediction model architecture with limited mature cases; (3) Systematically supporting channel growth.

The system architecture constructs user tags from bill and app data. Bill data provides large‑scale consumption records, while app data offers high‑value behavioral signals. A three‑layer tokenization system includes vectorization (character‑level and word‑level), rule‑based merging, and a probabilistic graph inference model to filter noisy or fraudulent bills.

Example: a bill "Alipay consumption – Yang Guo Fu Spicy Hotpot – **amount**" is split into key entities and other parts; the probabilistic graph captures co‑occurrence probabilities to assess phrase strength and filter out irrelevant or risky entries.

App‑based tagging uses one‑vector (single dimension) and cross‑vector (2‑D or 3‑D) methods, incorporating description fields to enrich feature representation and enable predictions such as gender, age, or financial interest categories.

Model deployment pipeline: source code in Python, C++, or R is translated to a stable runtime environment; feature generation includes real‑time scanning, offline auto‑integration, and multi‑source fusion; training supports online and offline modes; evaluation metrics cover AUC, Recall, Precision, F1‑score, MAP, K‑fold volatility; robustness testing involves simulated online tests and calibration. The 51CG operation platform automates the entire workflow, enabling rapid model rollout.

Growth case 1: Personalized A/B testing for new‑user zones (e.g., onboarding, repayment, investment, card loan, credit) achieved a 7‑14% conversion lift over default ranking.

Growth case 2: Push optimization focuses on five questions—timing, copy, audience, content, fatigue control. Timing is refined via an A/B iteration based on active‑hour distributions. Copy selection uses a Multi‑Armed Bandit (UCB) algorithm to balance exploration and exploitation.

Push open‑rate prediction combines rule‑based recent‑N‑day features, a Logistic Regression model, and a calibrated fusion to limit over‑frequency and improve ROI.

Growth case 2 (continued): User segmentation (low‑activity churn, high‑activity churn, cyclical, trial) drives differentiated push strategies, yielding significant ROI improvements across all groups after a month of A/B testing.

Author: Chen Bingqiang, Senior Algorithm Expert at 51 Credit Card Manager, former Director of Algorithm R&D at US‑based Nanjee and head of ID‑Graph team at a leading Chinese big‑data company, with extensive experience in financial ML/DL research and deployment.

Community: DataFun, a practical data‑science community offering offline deep‑dive salons and online content, aims to share industrial expertise with big‑data and AI practitioners.