A Framework using SVM based Rule Extraction on Derived Feature Set for High Accuracy Classification in Real Time Detection Systems

摘要

The effectiveness of detection engines used to predict potential anomalies in real-time data are limited by precision of detection rules which are predominantly based on domain knowledge. This work explores a framework that uses explicitly-defined relationships between data features in combination with embedded relationships revealed by time-based analysis. The framework utilizes machine learning techniques such as Evolutionary Computing, Support Vector Machines, and Decision Trees to produce highly accurate classification rules for real-time detection system. Real-time credit card fraud data was used to test and validate the framework.