Multidimensional Data Clustering and Visualization Using Hybrid Artificial Neural Network

All data with higher dimensions than 3D require some form of transformation to reveal its potential hidden patterns to assist human perception in making better data-related decisions. The exponential growth of multidimensional unlabeled data generated each day makes unassisted human perception harder to harness the full potential of the collected data that would provide insights. In order to bridge the gap between human visual perception and the multidimensional nature of modern-day data, this paper proposed a unique way of multidimensional data clusters discovery through a hybrid Artificial Neural Network (ANN) algorithms of Self-Organizing Map and modified Adaptive Coordinate (SOM-mAC). A Series of experiments conducted on synthetic 3D Gaussians dataset and benchmarking datasets Iris, Wine, and Wisconsin Breast Cancer (WBC) demonstrated that the proposed SOM-mAC algorithm could identify, cluster, and label the dataset's underlying topologies and densities. The algorithm can also identify and isolate potential outliers associated with each cluster. As clustering can be done in 2D through single-passing of the dataset through trained network, this algorithm has the potential for faster clustering of high dimensional big data.

I. Introduction

Multidimensional data are data with more than three attributes. However, human visual perception is limited to three dimensions. Therefore, multidimensional data must be transformed into three or fewer dimensions so that their potential hidden patterns can be revealed to assist humans in making better data-related decisions. The progression of surveys conducted by Chan [1], Liu et al. [2], and Espadoto et al. [3], and Espadoto et al. [4] gave a good overview of the different transformation techniques for multidimensional data exploration. From the data perspective, these techniques can be broadly categorized into labeled or unlabeled data. This research focused on the unlabeled data category, which is more relevant to the Big Data analytic paradigm.