Automatic clustering based on crow search algorithm-kmeans (CSA-kmeans) and data envelopment analysis (DEA)

Abstract

Cluster Validity Indices (CVI) evaluate the efficiency of a clustering algorithm and Data Envelopment Analysis (DEA) evaluate the efficiency of Decision-Making Units (DMUs) using a number of inputs data and outputs data. Combination of the CVI and DEA inspired the development of a new automatic clustering algorithm called Automatic Clustering Based on Data Envelopment Analysis (ACDEA). ACDEA is able to determine the optimal number of clusters in four main steps. In the first step, a new clustering algorithm called CSA-Kmeans is introduced. In this algorithm, clustering is performed by the Crow Search Algorithm (CSA), in which the K-means algorithm generates the initial centers of the clusters. In the second step, the clustering of data is performed from k min cluster to k max cluster, using CSA-Kmeans. At each iteration of clustering, using correct data labels, Within-Group Scatter (WGS) index, Between-Group Scatter (BGS) index, Dunn Index (DI), the Calinski-Harabasz (CH) index, and the Silhouette index (SI) are extracted and stored, which ultimately these indices make a matrix that the columns of this matrix indicate the values of validity indices and the rows or DMUs represent the number of clustering times from k min cluster to k max cluster. In the third step, the efficiency of the DMUs is calculated using the DEA method based on the second stage matrix, and given that the DI, CH, and SI estimate the relationship within group scatter and between group scatter, WGS and BGS are used as input variables and the indices of DI, CH and SI are used as output variables to DEA. Finally, in step four, AP method is used to calculate the efficiency of DMUs, so that an efficiency value is obtained for each DMU that maximum efficiency represents the optimal number of clusters. In this study, three categories of data are used to measure the efficiency of the ACDEA algorithm. Also, the efficiency of ACDEA is compared with the DCPSO, GCUK and ACDE algorithms. According to the results, there is a positive significant relationship between input CVI and output CVI in data envelopment analysis, and the optimal number of clusters is achieved for many cases.