Discrete Nonnegative Spectral Clustering

摘要: Spectral clustering has been playing a vital role in various research areas. Most traditional spectral algorithms comprise two independent stages (e.g., first learning continuous labels and then rounding the learned into discrete ones), which may cause unpredictable deviation of resultant cluster from genuine ones, thereby leading to severe information loss performance degradation. In this work, we study how achieve as well reliably generalize unseen data. We propose novel scheme deeply explores label properties, including discreteness, nonnegativity, discrimination, learns robust out-of-sample prediction functions. Specifically, explicitly enforce transformation on intermediate labels, leads tractable optimization problem with solution. Besides, preserve natural nonnegative characteristic enhance interpretability results. Moreover, further compensate unreliability integrate an adaptive module $\ell _{2,p}$ learn function for grouping also show that component can inject discriminative knowledge under certain conditions. Extensive experiments conducted data sets have demonstrated superiority our proposal compared several existing approaches.