Rule Learning from Time-Dependent Data Applied to Fraud Detection
Abstract
In financial environment, fraud detection is a challenging problem with tremendous financial impacts where data is highly unbalanced, sequential and timestamped. An additional constraint comes from the fact that common machine learning methods cannot be used alone for fraud detection, as every decision made in order to label a transaction as fraudulent needs to be explainable and the complete model understandable.The use of a symbolic language, such as understandable classification rules, is therefore preferred or even required.
Origin : Files produced by the author(s)