Generalized Deep Embedded Fuzzy C-Means for Clustering High-Dimensional Data

Clustering is one of the fundamental techniques of machine learning. Its integration with deep neural networks allows for extracting robust feature representations and yielding better clustering results. Deep-embedded clustering algorithms employ external information to form auxiliary and target distributions minimized via KL divergence. This study introduces a Generalized Deep Embedded Fuzzy C-Means (GDeeFCM) algorithm that learns both feature representations and cluster assignments at the same time. The principal advantage of GDeeFCM is using the objective function of the clustering algorithm, FCM in our case, as the loss function to update the encoder weights and clusters center simultaneously without the need to adapt information from external sources. It is worth mentioning that FCM is selected for this study, but any clustering algorithm can be employed. Our model's performance is compared with similar structures that utilize t-SNE for soft label assignments and KL Divergence as the loss function. Experimental results on four image datasets demonstrate the effectiveness of our algorithm.