Eye-Clustering: An Enhanced Centroids Prediction for K-means Algorithm

Abstract

Unsupervised machine learning is a powerful technique for performing clustering, which involves identifying patterns or similarities within a dataset and grouping them into distinct clusters or subgroups. Various clustering methods are available, including K-means, hierarchical clustering, and density-based clustering. Among these, K-means is widely used for efficiently solving clustering problems. This work aims to enhance the performance of the K-means algorithm by introducing a novel method for selecting the initial centroids, thereby minimizing randomness and reducing the number of iterations needed to reach optimal results. The proposed method, named Eye-means, emulates the natural ocular process of estimating initial centroids. To achieve this goal, supervised machine learning was employed to train models on graphs with labeled data points, where each graph contains a set of points and a label indicating the centroid determined by K-means. Hundreds of such labeled graphs were used to train the model to predict the location of centroids. The objective is to produce a model capable of predicting centroids with greater accuracy than the traditional random initialization used in K-means. Experimental results indicate that the proposed method outperforms random initialization in terms of the number of iterations required to achieve an optimal solution.