Fanglue: An Interactive System for Decision Rule Crafting in Fraud Detection

In adversarial security and risk‑control scenarios, rapidly evolving fraud tactics often render training data insufficient for algorithms to automatically discover high‑quality interception rules, necessitating the incorporation of expert domain knowledge. Fanglue, an interactive rule‑development system built by Ant Group AI Infra, addresses this challenge by providing a web interface that visualizes rule creation and continuously recommends candidate conditions with associated evaluation metrics.

The system architecture consists of three core front‑end modules—Condition Recommendation, Condition Editing, and Rule Evaluation—backed by a Ray‑based distributed computation engine. Data are horizontally partitioned and loaded into a pool of Ray Actors; each Actor computes local statistics, which are aggregated by a Driver and returned to the Task Manager for further processing and display.

Technically, Fanglue adopts a conjunctive normal form (CNF) representation supporting both AND and OR conditions, and offers three real‑time recommendation types: AND‑condition, OR‑condition, and Approximate‑condition. Candidate triples (feature, operator, value) are exhaustively evaluated using distributed actors, and top‑ranking candidates are presented to users for selection, enabling rapid exploration of the rule space.

Experimental results on a 1.4‑million‑sample, 50‑feature dataset demonstrate that Fanglue can generate AND‑candidate recommendations within one second using 16 Ray actors, satisfying interactive latency requirements. The platform also supports sequential coverage mode, allowing users to iteratively edit multiple rules while excluding already covered data.

Fanglue has been deployed in Ant Group’s internal security risk‑control pipelines and shared with external financial institutions, improving rule‑development efficiency, reducing costs, and enhancing rule accuracy. Future work will focus on optimizing recommendation algorithms and evaluation metrics to handle more complex, dynamic scenarios.