model share subspac boost multi label classif rong yan jelena tesic john smith typic approach multi label classif problem requir learn independ classifi label exampl featur comput bottleneck sizeabl dataset larg label space paper propos effici effect multi label learn algorithm call model share subspac boost mssboost attempt reduc inform redund learn process algorithm automat find share combin number base model multipl label model learn random featur subspac boot trap data sampl decis function label jointli estim small number share subspac model support entir label space experiment result synthet data real multimedia collect demonstr propos algorithm achiev classif perform ensembl baselineclassifi signific speedup learn predict process smaller number base model achiev classif perform model share counterpart 