MoRE: Multi-output residual embedding for multi-label classification

Highlights：

• We further consider label correlations between instances and their neighbors to get more appropriate distance metric. Therefore, our proposed approach learns residual projections not only from instance residuals, but also from multiple residuals of its neighbors. Through this, we learn a better low-rank projection in label space.

• We conduct comprehensive comparative studies over eight benchmark data sets with diverse characteristics to validate the effectiveness of our proposed approach. Experimental results obtained from multiple evaluation metrics validate that MoRE has better predictive performance than many existing state-of-the-art MLC methods.

摘要

•Different from most existing methods that explore relationships between features and labels directly, we learn a low-rank projection in the label space by analyzing the residual structure between the input and output spaces. The proposed method then encodes residual projections to embed data from label space to a lowerdimensional space. In this way, our proposed approach provides a reasonable distance metric algorithm for MLC to achieve dimension reduction in label space.•We further consider label correlations between instances and their neighbors to get more appropriate distance metric. Therefore, our proposed approach learns residual projections not only from instance residuals, but also from multiple residuals of its neighbors. Through this, we learn a better low-rank projection in label space.•We conduct comprehensive comparative studies over eight benchmark data sets with diverse characteristics to validate the effectiveness of our proposed approach. Experimental results obtained from multiple evaluation metrics validate that MoRE has better predictive performance than many existing state-of-the-art MLC methods.