Title: Deep clustering: A new clustering method in the sequential approach
Authors: Claudia Rampichini - University of Rome La Sapienza (Italy) [presenting]
Abstract: Deep clustering is a recent technique that exploits the potential of neural networks to overcome the problems of conventional clustering methods. Specifically, an autoencoder neural network is used to obtain a dimensional reduction in which a clustering algorithm is involved; the approach can be sequential or simultaneous. The focus is on a new clustering method that uses and enhances the information provided by the membership degrees, derived from a fuzzy algorithm, in order to improve clustering performance. First, a fuzzy algorithm is applied, and then the units that have an unclear assignment are reclassified using a crisp algorithm. A unit has an unclear assignment when membership degrees are close to each other. A summary measure of membership degrees is chosen and, based on a threshold value, a subset of the units are reclassified. The adequacy of the proposal is checked by means of several benchmark data sets and the results show margins for improvement in performance compared to both fuzzy and crisp algorithms.