Semi-Supervised Kernel Mean Shift Clustering

Abstract

Abstract-Mean shift clustering is a powerful nonparametric technique that does not require prior knowledge of the number of clusters and does not constrain the shape of the clusters. However, being completely unsupervised, its performance suffers when the original distance metric fails to capture the underlying cluster structure. Despite recent advances in semi-supervised clustering methods, there has been little effort towards incorporating supervision into mean shift. We propose a semi-supervised framework for kernel mean shift clustering (SKMS) that uses only pairwise constraints to guide the clustering procedure. The points are first mapped to a high-dimensional kernel space where the constraints are imposed by a linear transformation of the mapped points. This is achieved by modifying the initial kernel matrix by minimizing a log det divergence-based objective function. We show the advantages of SKMS by evaluating its performance on various synthetic and real datasets while comparing with state-of-the-art semi-supervised clustering algorithms.

Constrained Low-Rank Representation for Robust Subspace Clustering

“…Therefore, it would be crucial to incorporate these information for constructing discriminative affinity matrix. Although, semisupervised learning approaches [2], [12], [28], [33], [45], [46], [50], [53] have received a great attention recently, few have utilized semisupervised representation-based methods. Existing methods extend an unsupervised learning to a semisupervised setting usually by graph based regularization.…”

Enhancing Hyperspectral Endmember Extraction Using Clustering and Oversegmentation-Based Preprocessing

“…For this purpose, we perform three well-known clustering methods i.e., k-means [31], fuzzy c-means (FCM) [32], and a semisupervised kernel mean shift clustering (SKMS) [33] to evaluate their overall performances on COPP. Prior to developing three clustering procedures, principal components transformation (PCT) is performed on original hyperspectral dataset [34].…”

“…FCM provides the best result for overlapped dataset and comparatively better than k-means algorithm. In the case of SKMS, just three PCs are utilized as the result of its computational complexity and high execution time (using C compiler) [33]. In order to decrease computational load, we develop MATLAB super-pixel implementation available online 1 on hyperspectral image prior SKMS.…”

“…To provide pairwise must-link and cannot-link constraints, 10 labeled pixels per cluster are randomly adopted. Details of SKMS procedure are presented in [33]. By SKMS, the image pixels are first projected to a high-dimensional kernel space where the constraints are applied by a linear transformation of the projected pixels.…”

“…By SKMS, the image pixels are first projected to a high-dimensional kernel space where the constraints are applied by a linear transformation of the projected pixels. In SKMS, the pairwise must-link constraints are exploited to automatically adopt the bandwidth parameter [33].…”