Integrated Rough Fuzzy Clustering for Categorical data Analysis

Abstract In recent times, advanced data mining research has been mostly focusing on clustering of categorical data, where a natural ordering in attribute values is missing. To address this fact the Rough Fuzzy K-Modes clustering technique has been recently developed in order to handle imperfect information, i.e. indiscernibility (coarseness) and vagueness within the dataset. However, it has been observed that the said technique suffers from the problem of local optima due to the random choice of initial cluster modes. Hence, in this paper, we have proposed an integrated clustering technique using multi-phase learning. In this regard, first, Simulated Annealing based Rough Fuzzy K-Modes and Genetic Algorithm based Rough Fuzzy K-Modes are proposed in order to perform the clustering better by considering clustering as an underlying optimization problem. These clustering methods individually produce clusters having set of central and peripheral points. Thereafter, for each case, final improved clustering results are obtained by assigning peripheral points to a particular crisp cluster using Random Forest, where central points are used as training set. Second, the varying cardinality of the training and testing sets produced by each clustering method further motivated us to propose a generalized technique called Integrated Rough Fuzzy Clustering using Random Forest, where, results of three aforementioned clustering techniques are used to compute the roughness measure. Based on this measure, three different sets namely best central points, semi-best central points and pure peripheral points are determined. Thereafter, using multi-phase learning, best central points are used to classify the semi-best central points and then using both of them, pure peripheral points are classified by Random Forest. Experimental results are reported quantitatively and visually to demonstrate the effectiveness of the proposed methods in comparison with well-known state-of-the-art methods for six synthetic and five real-life datasets. Finally, statistical significance tests are conducted to establish the superiority of the results produced by the proposed methods.