Revealing Abnormality Based on Hybrid Clustering and Classification Approach

Abstract

Abnormality Detection is the process of locating abnormal instances within the data. In this work, we have applied Abnormality Detection to the domain of detection associated with Credit Card Fraud. This problem is actually attributed to demonstrating those credit card transactions which have occurred in the earlier times, with the presence of awareness related to those instances, which are actually fraud ones. Applying this model, we can use it to predict if a new transaction is a fraud based or not. In this proposed work, we have utilized a combination framework of data mining clustering algorithms so as to solve the problem of credit card fraud detection to a particular extent. The proposed work Revealing Abnormality Based on Hybrid Clustering and Classification Approach (RA-HC-CA) consists of two stages namely a clustering phase followed by a detection phase. In the clustering phase, we have employed a combined clustering approach initiated by k-means clustering algorithm followed by hierarchical clustering algorithm. Prior to Hierarchical/Agglomerative clustering, the whole data set is clustered into meaningful ‘k’ knots by k-means clustering procedure. The output of ‘k’ groups is then inputted to Agglomerative clustering algorithm to merge the already obtained ‘k’ clusters from the previous phase, into more meaningful clusters. This is continued until 70–75% of data falls on one large group, which is the Normal group. The remaining data instances may converge in various other abnormal groups. The strong assumption made here is that such clusters with less instances, than a particular threshold are considered to be groups pertaining to fraud ones. Then, so as to check for the presence of an instance as fraud one, we initially identify the proximate gathering to which it fit into. Then, within that identified cluster, LDA (Linear Discriminant Analysis) is carried out. It has been observed that the proposed approach (RA-HC-CA) achieved 80.5% accuracy in comparison with various other existing methods.