Optimization of the Numeric and Categorical Attribute Weights in KAMILA Mixed Data Clustering Algorithm

Abstract

The mixed data clustering algorithms have been timidly emerging since the end of the last century. One of the last algorithms proposed for this data-type has been KAMILA (KAy-means for MIxed LArge data) algorithm. While the KAMILA has outperformed the previous mixed data algorithms results, it has some gaps. Among them is the definition of numerical and categorical variable weights, which is a user-defined parameter or, by default, equal to one for all features. Hence, we propose an optimization algorithm called Biased Random-Key Genetic Algorithm for Features Weighting (BRKGAFW) to accomplish the weighting of the numerical and categorical variables in the KAMILA algorithm. The experiment relied on six real-world mixed data sets and two baseline algorithms to perform the comparison, which are the KAMILA with default weight definition, and the KAMILA with weight definition done by the traditional genetic algorithm. The results have revealed the proposed algorithm overperformed the baseline algorithms results in all data sets.