From Chaotic Transaction Data to Intelligent Insight: AI Raises Classification Accuracy by 20%

Background and Challenge

Amid rapid digital transformation, electronic payment systems generate massive, unstructured transaction logs. Traditional classification methods—merchant category codes (MCC) and keyword matching—suffer from source distortion and rule limitations, leading to inaccurate user profiles, weak risk control, and ineffective marketing.

Solution Architecture

The R&D team designed an end‑to‑end intelligent classification engine centered on natural‑language processing. A BERT model serves as the core, enhanced by a multi‑stage training strategy and knowledge‑distillation to meet high‑throughput, low‑latency requirements.

Data Foundations

Two proprietary datasets were created: (1) an unsupervised pre‑training corpus built from tens of millions of cleaned merchant descriptions, capturing financial‑domain language patterns; (2) a supervised fine‑tuning set covering dozens of sub‑scenarios, produced via a human‑in‑the‑loop labeling workflow that combines expert annotation with iterative model feedback.

Three‑Stage Training Strategy

The model follows a progression: general pre‑training → domain‑specific pre‑training → task‑specific fine‑tuning. This hierarchy first equips the model with broad linguistic knowledge, then deepens its understanding of banking terminology, and finally optimizes it for the transaction‑type classification task. The architecture relies on the Transformer‑based BERT with self‑attention to capture deep semantic relations in merchant descriptions.

Performance Optimization via Knowledge Distillation

A teacher‑student framework was employed: a large ensemble of BERT models formed the teacher, providing high‑accuracy predictions; a lightweight student model learned both the ground‑truth labels and the teacher’s soft outputs. This distillation preserved accuracy while dramatically reducing inference latency and GPU memory consumption, enabling scalable deployment.

Results

Offline validation showed the intelligent model achieved an average accuracy above 95%, roughly a 20‑percentage‑point improvement over the legacy rule‑based system, with coverage exceeding 90% for new merchant types. The refined classification split previously coarse categories (e.g., generic “shopping”) into detailed scenes such as “dining”, “entertainment”, and “education”.

Business Impact

The granular insights empowered business units to conduct deeper structural analyses, identify growth‑potential segments (e.g., “fund investment” during payment pressure periods), and construct a data‑driven “marketing calendar” that aligns campaigns with seasonal transaction trends across education, travel, and shopping domains.

Future Outlook

The semantic‑centric framework is positioned as a reusable foundation for downstream banking applications, including enhanced customer profiling and decision‑support systems, steering the institution toward a continuously learning, AI‑enabled digital bank.