An incomplete multi-view multi-label learning with Universum

Abstract In current era, the incomplete multi-view multi-label data sets are widely encountered and they are hard to be processed due to the structure of each instance is complicated and some useful information maybe lost. To this end, this study uses the missing-information-index matrices and adopts Universum learning to promote the enhancements of useful information. Then we combine the self-adaptive measurements of within-view correlation, cross-view correlation, consensus-view correlation among features, labels, and instances in together so that the missing information can be recovered with a more feasible way. The developed learning machine is named as incomplete multi-view multi-label learning with Universum (IMVML-U). In order to demonstrate the superiority of IMVML-U, we adopt \(5\) MVML data sets for experiments and according to the experimental results, it is found that IMVML-U can process incomplete multi-view multi-label data sets much better, update the correlations among features, labels, and instances dynamically, and enhances the useful information with a less time complexity increased.