Anomaly detection based multi label classification using Association Rule Mining (ADMLCAR)

Abstract

Multi label classification contains multiple labels in the label space. Any Multi label classification problem (MLC) deals with numerous class labels associated with data instances. Due to this, correct prediction of labels for a test data remains as a challenge in this field. In this paper an Anomaly Detection based Multi Label Classification using Association Rule Mining (ADMLCAR) is used for solving MLC problem. Conventionally, most of the multi label classification problem is solved by either of the two methods: Problem transformation, Algorithm adaptation. But the method discussed in this paper aims at a novel method different from traditional solution to multi label classification problem. For clustering, ADMLCAR uses k-means algorithm and for association rule mining purpose it uses vertical data format. To predict the test data instances, this method locates the nearest cluster. Once the clusters are identified it uses oversampling principal component analysis (PCA) within the nearest cluster with respect to test instances. Oversampling PCA is used to emphasize the need for confirming the fact that test instance's label set will not only be confined to its nearest cluster label set. This is because, anyways the test instance will be identified to a nearest cluster by means of Euclidean distance measure but as clustering is unsupervised the nearest cluster may contain many objects entities of different label sets.