Data Clustering Method based on Mixed Similarity Measures

Abstract

Data clustering aims to organize data and concisely summarize it according to cluster prototypes. There are different types of data (e.g., ordinal, nominal, binary, continuous), and each has an appropriate similarity measure. However when dealing with mixed data set (i.e., a dataset that contains at least two types of data.), clustering methods use a unified similarity measure. In this study, we propose a novel clustering method for mixed datasets. The proposed mixed similarity measure (MSM) method uses a specific similarity measure for each type of data attribute. When computing distances and updating clusters’ centers, the MSM method merges between the advantages of k-modes and K-means algorithms. The proposed MSM method is tested using benchmark real life datasets obtained from the UCI Machine Learning Repository. The MSM method performance is compared against other similarity methods whether in a non-evolutionary clustering setting or an evolutionary clustering setting (using differential evolution). Based on the experimental results, the MSM method proved its efficiency in dealing with mixed datasets, and achieved significant improvement in the clustering performance in 80% of the tested datasets in the non-evolutionary clustering setting and in 90% of the tested datasets in the evolutionary clustering setting. The time and space complexity of our proposed method is analyzed, and the comparison with the other methods demonstrates the effectiveness of our method.