Addressing Multi-Label Learning with Missing Labels via Feature Relevance guided Scaled Model Coefficients

Feature relevance is pivotal in multi-label learning tasks, as certain features have a greater impact on labels than others. Given the high dimensionality of multi-label datasets, efficiently identifying and utilizing feature relevance in model coefficients determination is essential. This article introduces a multi-label learning with missing labels approach using a modified -norm regularization to scale model coefficients based on feature relevance and ensure global sparsity similar to the Frobenius norm. The approach also incorporates label correlations learning for missing label recovery and expresses model coefficients through learned correlation decomposition. A squared Hinge loss function, known for its robust classification properties, is employed as an empirical loss measure. Experimental validation against six existing multi-label learning methods across six datasets and five metrics demonstrates the efficacy of our unified approach in both recovering missing labels and predicting new instance labels.

Recommended citation: Kumar, Sanjay, and Reshma Rastogi. "Addressing Multi-Label Learning with Missing Labels via Feature Relevance guided Scaled Model Coefficients." In International Conference on Pattern Recognition, Springer Nature Switzerlan, 2024