Deep evidential clustering based on feature representation learning and belief function theory

Evidential clustering is a promising clustering framework based on belief function theory which generalizes hard, fuzzy, possibilistic, and rough clustering. It allows us to gain deeper insight into the structure of the data. However, traditional evidential clustering algorithms cannot deal well with high-dimensional raw data such as text and images by the way of clustering on the original data representations directly. To address this problem, a deep evidential clustering (DEC) algorithm based on feature representation learning and belief function theory is proposed in this study. This algorithm incorporates the evidential clustering loss, the autoencoder reconstruction loss, and the regularization loss of network parameters to construct a combined optimization model. DEC performs deep feature extraction and evidential clustering simultaneously to ensure that the learned deep features are the evidential clustering-friendly representations of the raw data. In addition, DEC introduces evidential partition so that an object can belong to either a single class or any subset of a collection of classes, which provides more powerful expressive ability for uncertain data. Extensive experiments were conducted on real high-dimensional datasets, and the experimental results illustrated the superiority of DEC compared to the state-of-the-art evidential clustering and deep clustering algorithms.