A new preference disaggregation method for clustering problem: DISclustering

Abstract

Clustering, a famous technique in data analysis and data mining, attempts to find valuable patterns in datasets. In this technique, a set of alternatives is partitioned into logical groups which are called clusters. The partitioning is based on some predefined attributes to find clusters in which their alternatives are similar to each other comparing to other clusters. In conventional methods, the similarity is usually defined by a distance-based measurement, whereas in this study, we have proposed a new multi-attribute preference disaggregation method called DISclustering in which a new measurement named global utility is introduced for cluster similarity. In DISclustering, the global utility of each alternative is calculated through a feed-forward neural network in which its parameters are determined using SA algorithm. Each alternative is assigned to a cluster based on comparing the obtained global utility with cluster boundaries, called utility thresholds; aim to minimize the intra-cluster distances (ICD). For this purpose, all utility thresholds are estimated using PSO algorithm. The performance of the proposed method is compared with 18 clustering algorithms on 14 real datasets based on F-measure and object function values (ICD values using intra-cluster or Gower distances). The experimental results and hypothesis statistical test indicate that DISclustering algorithm significantly improved clustering results on F-measure criteria in which outperforms in almost 13 compared algorithms out of 18. Note that, DISclustering calculates cluster centroid in a different way comparing to other algorithms. Hence, its ICD values are less eligible to perform a fair comparison.