Unified Expression Ripple Down Rules based Fraud Detection Technique for Scalable Data

Fraud detection for online banking is an important research area and higher accuracy is highly desirable. The main challenges in fraud analysis are due to the presence of heterogeneous transactions data, large and distributed data. Among existing rule-based techniques for fraud detection, Ripple Down Rules (RDR) is ideal due to its less maintenance and incremental learning. However, banking data sets contains billions of transactions, so the performance of RDR on distributed and Big data platforms need to be studied for fraud detection applications. A Unified Expression RDR fraud deduction technique for Big data has been proposed and evaluated in this paper. By incorporating the Unified Expressions into the RDR and evaluating the expressions using the Lift score, the compactness of the ruleset can be achieved and the accuracy of the classification improved. In addition, the paper presents a compact model that fuses Majority and Minority classes for RDRbased classifiers. Classification accuracy is compared with the two existing RDR implementations RIDOR and Integrated Prudence Analysis technique and a non-RDR classifier as well. Empirical evaluations on various datasets have shown that not only the ruleset size of training and prediction dataset is reduced, but the accuracy of classification is also improved. The results showed an improvement in the classification accuracy when compared to two RDR and non-RDR based classifiers. The proposed technique is used for experimental validation and the development of fraud analysis, but it can also be used in other domains, in particular for scalable and distributed systems.