PROJECT TITLE :

Discrete Nonnegative Spectral Clustering - 2017

ABSTRACT:

Spectral clustering has been enjoying a vital role in various analysis areas. Most ancient spectral clustering algorithms comprise two independent stages (e.g., 1st learning continuous labels and then rounding the learned labels into discrete ones), that might cause unpredictable deviation of resultant cluster labels from real ones, thereby resulting in severe info loss and performance degradation. In this work, we tend to study how to realize discrete clustering and reliably generalize to unseen data. We have a tendency to propose a completely unique spectral clustering theme that deeply explores cluster label properties, as well as discreteness, nonnegativity, and discrimination, also learns robust out-of-sample prediction functions. Specifically, we tend to explicitly enforce a discrete transformation on the intermediate continuous labels, which leads to a tractable optimization problem with a discrete solution. Besides, we have a tendency to preserve the natural nonnegative characteristic of the clustering labels to boost the interpretability of the results. Moreover, to further compensate the unreliability of the learned clustering labels, we integrate an adaptive robust module with l2p loss to be told prediction operate for grouping unseen information. We additionally show that the out-of-sample element will inject discriminative data into the training of cluster labels underneath certain conditions. Intensive experiments conducted on varied data sets have demonstrated the superiority of our proposal as compared to several existing clustering approaches.