Detecting Anomaly Using Partitioning Clustering with Merging

Nowadays, anomaly detection in data is an important field of research. A couple of methods have already been developed for this purpose, among which clustering-based anomaly detection remains an important one. Most methods using clustering approaches consider either numeric or categorical attributes of the data instance, but not both. However, in real life, most datasets are hybrids containing numeric, categorical, or mixed data. For such hybrid datasets, the typical distance formula does not work. This paper presents a unified distance formula for the clustering approach, in which the attributes may be numeric, categorical, or a mixture of both. Here, we propose an anomaly detection method based on both partitioning and hierarchical approaches. The algorithm proposed here is a modified version of k-means clustering algorithm in the sense that it also uses a suitably defined merge function. After every iteration the merge function is used to merge a pair of highly similar clusters to produce a larger cluster. The efficacy of our method is established using the complexity analysis and experimental results