Local search genetic algorithm-based possibilistic weighted fuzzy c-means for clustering mixed numerical and categorical data

Clustering for mixed numerical and categorical attributes has attracted many researchers due to its necessity in many real-world applications. One crucial issue concerned in clustering mixed data is to select an appropriate distance metric for each attribute type. Besides, some current clustering methods are sensitive to the initial solutions and easily trap into a locally optimal solution. Thus, this study proposes a local search genetic algorithm-based possibilistic weighted fuzzy c-means (LSGA-PWFCM) for clustering mixed numerical and categorical data. The possibilistic weighted fuzzy c-means (PWFCM) is firstly proposed in which the object-cluster similarity measure is employed to calculate the distance between two mixed-attribute objects. Besides, each attribute is placed a different important role by calculating its corresponding weight in the PWFCM procedure. Thereafter, GA is used to find a set of optimal parameters and the initial clustering centroids for the PFCM algorithm. To avoid local optimal solution, local search-based variable neighborhoods are embedded in the GA procedure. The proposed LSGA-PWFCM algorithm is compared with other benchmark algorithms based on some public datasets in UCI machine learning repository to evaluate its performance. Two clustering validation indices are used, i.e., clustering accuracy and Rand index. The experimental results show that the proposed LSGA-PWFCM outperforms other algorithms on most of the tested datasets.