Abstract
In real-world scenarios, high-dimensional data are often associated with diverse class labels, hindering the efficacy of the data mining process. Accordingly, a wave of investigations into multi-label feature selection has been sparked to confront this major challenge. Nevertheless, prior studies have paid scant attention to class skewness, an intrinsic property of multi-label data, manifested across labels. To this end, this paper proposes a feature selection methodology tailored for these class-imbalance multi-label datasets. Specifically, we embed graded weights into the penalty term to mitigate deviations from the underlying label structure and model bias towards common labels. As a complement, we further endow the feature manifold, which is constrained by the latent label space, with instance-specific graph contributions to better preserve the local structure involving rare classes. Furthermore, we explicitly impose dynamic label graph regularization on feature weights. Finally, we incorporate all aforementioned terms into a linear mapping learning paradigm, with a trustworthy solution utilizing alternating iterations devised for optimizing the objective function. Extensive experiments on 14 multi-label benchmarks with varying degrees of class skewness demonstrate the proposed method’s effective identification of informative features.