A novel data visualization method for the effective assessment of cluster tendency through the dark blocks image pattern analysis

Abstract

Data clustering is the most promising unsupervised technique, and it can be used to partition the data objects into different clusters. The data objects are placed according to derived similarity features using distance metrics, such as Euclidean, cosine, etc. The top clustering algorithms include k-means, and hierarchical clustering algorithms are extensively used and succeeded in many real-life applications. These algorithms are suffered from cluster tendency problems. They detect the number of clusters before the unlabelled dataset is called cluster tendency. Various pre-cluster tendency methods are studied, and it found that visual assessment of cluster tendency (VAT) solves the issue of cluster tendency. VAT cannot effectively derive the number of clusters for compact-separated (CS) datasets and obtain fewer groups for non-compact separated (non-CS) datasets. The Spectral-based implementations are recommended for the best assessment of cluster tendency, especially for the type of non-CS datasets. With this motivation, current data visualization methods are enhanced with spectral modeling to overcome the problem of cluster tendency over the non-CS datasets, which recognizes the cluster patterns efficiently with the spectral features of non-CS datasets. Comparative demonstrations were presented for the experimental study of existing and proposed data visualization methods using various performance parameters.