Integrating Interval Type-2 Fuzzy Sets into Deep Embedding Clustering to Cope with Uncertainty

Working with unlabeled data carries the burden of uncertainties especially when the data are high-dimensional. Clustering is not an exception in this aspect and it requires special treatment. In this study, we propose to cope with the uncertainties which occur during clustering high-dimensional data with Interval Type-2 (IT2) Fuzzy Sets (FSs) and Deep Learning (DL) methods. Generation of the IT2-FSs is done with different cluster similarity functions parameterized with Interval Valued Parameters (IVPs). These parameters are introduced as the representations of the uncertainty in cluster assignments. As the backbone of the proposed method, Deep Embedding Clustering (DEC) is employed. The resulting IT2 fuzzy clustering inference is integrated into DEC so that both the inference and the training of the proposed model are operational in popular DL frameworks. Therefore, for a straightforward deployment, the constraints on IT2-FSs are redefined by introducing parameterization tricks upon IVPs. The presented comparative results indicate that coping with the uncertainties through IT2-FSs is superior to their baseline type-1 counterparts.

Automated summary

This study proposes to address uncertainties in high-dimensional data clustering using Interval Type-2 Fuzzy Sets and Deep Learning methods. The Interval Type-2 Fuzzy Sets are generated with different cluster similarity functions parameterized with Interval Valued Parameters, introducing representations of uncertainty in cluster assignments. The integration of IT2 fuzzy clustering inference into Deep Embedding Clustering shows superior results compared to baseline type-1 counterparts in coping with uncertainties.