High-Dimensional Data Clustering with Fuzzy C-Means: Problem, Reason, and Solution

Fuzzy C-Means (FCM) clustering algorithm is a popular unsupervised learning approach that has been extensively utilized in various domains. However, in this study, we point out a major problem faced by FCM when it is applied to the high-dimensional data, i.e., quite often the obtained prototypes (cluster centers) could not be distinguished with each other. Many studies have claimed that the concentration of the distance (CoD) could be a major reason for this phenomenon. This paper has therefore revisited this factor, and highlight that the CoD could not only lead to decreased performance, but sometimes also positively contribute to enhanced performance of the clustering algorithm. Instead, this paper point out the significance of features that are noisy and correlated, which could have a negative effect on FCM performance. Hence, to tackle the mentioned problem, we resort to a neural network model, i.e., the autoencoder, to reduce the dimensionality of the feature space while extracting features that are most informative. We conduct several experiments to show the validity of the proposed strategy for the FCM algorithm.