A Study on Efficient Clustering Techniques Involved in Dealing With Diverse Attribute Data

Introduction

In this web 3.0 era, information is accelerating from the Big data and with the applicability of IoT devices. Each day, information is generated from the use of Social sites to wearable devices and IOT’s. For instance, millions of google searchers, thousands of YouTube video uploads, data from service-based applications (like medical, transport, logistics, education, shopping sites, etc.), tweets, comments are being generated. With this information, researchers are trying to extract the patterns that help in analyzing and understanding data. However, this raw data cannot be analyzed using any algorithm. Often in real-time situations, data is not available with any appropriate classifications and pre-defined labels. Consequently, there is a need to develop certain models of machine learning capable of precise classification of the new data, based on certain similarities in features. This process can be accomplished by ‘Clustering’, an algorithm of unsupervised learning. In machine learning () as well as data mining, this analysis of clustering is a very important technique. The objective of clustering analysis is to segregate an ensemble of undefined objects into diverse clusters in such a way that the data objects of a specific cluster are either different or similar to the data objects of another cluster. The applications of cluster analysis are numerous including the categorization of customers, setting market targets, analysis of social networks, bioinformatics, and analysis of scientific data (Han & Kamber, 2000). Segmentation of a specific dataset into a homogeneous collection is performed by an optimization model of partitioning depicted by a cost function, in such a way that there is a similarity among the observations inside a cluster, while dissimilarity among the observations of other clusters. Input: An unlabeled training set with attribute values D={observations, i=1,…,N} with N objects described by d attributes where observations= {Attribute_1, Attribute_2,…, Attribute_d} ÎR^d K depicts the total number of initial clusters. Output: A set of K clusters C₁, C₂,…, C_k. The variations in size, shape, and density in the resultant clusters largely depend on the number of clusters K and the processes of clustering adopted. The prime characteristic feature of a good clustering is in its intense compactness, which means that the intra-cluster observations should be as proximate as can be possible, and isolation which means that the inter-cluster variations in observations should be as scattered as can be possible.

As illustrated in Figure 1, it can be noted that clusters C1 and C3 are different in shape but very compact, while C2 and C3 are comparatively not so compact. Certain observations are found to be secluded from the cluster’s core. Such secluded observations could be the representations of the outlier and noise in the resultant clusters and may not be causing a negative influence on the comprehension towards the close of the process (Ben Salem, Naouali, & Chtourou, 2018).

Figure 1.

Segmenting observations into groups based on inter and intra distance

d1 is the representation of the inter-distance existing between the observations associated with diverse clusters requiring to be maximized. This results in the procurement of isolated clusters. Similarly, d2 depicts the intra-distance existing between the observations pertaining to the same cluster requiring to be minimized. This results in the procurement of compact clusters.