<p>Stacked ensemble methods have demonstrated promising performance in multi-label classification tasks, but current implementations face two critical limitations that hinder the effectiveness. One is the conventional feature extraction approaches employed in these frameworks often fail to adequately capture the associations between feature and label, consequently yielding feature subspaces with suboptimal discriminative ability. The other is the error propagation phenomenon in stacked architecture leads to cumulative error amplification, where base classifier inaccuracies systematically propagate through meta-classifier training and subsequent prediction phases, ultimately resulting in performance decline of the entire ensemble system. To resolve these aforementioned problems, a novel multi-label stacked ensemble algorithm based on feature subspace extraction and label correction (MSSLC) is proposed. Firstly, a feature subspace extraction method built on feature similarity is presented, where the label correlations are measured and the feature weight matrix constraints are considered, this can obtain the optimized feature subspace. Secondly, the stacked ensemble classification based on label correction is developed. A label correction mechanism based on the label count range is adopted to reduce the output errors of meta-classifier, which can reduce the propagation and accumulation of errors in the meta-classifier’s training and prediction. Finally, comprehensive experimental validation is conducted. The MSSLC algorithm is assessed against twelve diverse comparison methods on twelve multi-label datasets. Extensive experiments confirm that the MSSLC attains superior results compared to the advanced algorithms.</p>