Multi-view　clustering,　which　improves　clustering　performance　by　exploring　complementarity　and　consistency　among　multiple　distinct　feature　sets,　is　attracting　more　and　more　researchers　due　to　its　wide　applications　in　various　fields　e.g.,　pattern　recognition　and　data　mining.　Traditional　approaches　usually　explore　above　characteristics　by　mapping　different　views　to　a　unified　embedding　through　view-specific　mapping　matrices　or　neural　networks.　Then　the　unified　embedding　is　fed　into　conventional　single　view　clustering　algorithms　for　final　clustering　results.　However,　a　unified　embedding　is　not　enough　to　model　distinct　or　even　conflict　multiple　view　characteristics　due　to　their　diverse　representation　abilities.　Moreover,　clustering　and　embedding　learning　are　divided　into　two　separate　parts,　which　may　bring　in　a　gap　between　the　class　label　and　the　learned　embedding.　To　alleviate　above　problems,　both　unified　and　view-specific　embeddings　are　learned,　and　a　shared　operator　tensor　and　view-specific　latent　variables　are　introduced　for　their　relationship　modeling.　Besides,　a　Kullback-Liebler　divergence　based　objective　is　developed　as　a　clustering　oriented　constraint,　which　leads　to　more　clustering　friendly　embedding　learned.　Extensive　experiments　are　conducted　on　six　widely　used　datasets,　achieving　better　results　compared　with　several　state-of-the-art　approaches.