Speed up kernel dependence maximization for multi-label feature extraction

Kernel dependence maximization for multi-label dimensionality reduction (kMDDM) has been proposed recently to cope with high-dimensional multi-label data. In order to produce discriminant projection vectors, kMDDM utilize the Hilbert-Schmidt independence criterion to capture the dependence between the feature description and the associated labels. However, the computation of kMDDM involves dense matrices eigen-decomposition that is known to be computationally expensive for large scale problems. In this paper, we reformulate the original kMDDM as a least-squares problem, so as to significantly lessen computational burden by utilizing the conjugate gradient algorithms. Further, appealing regularization techniques can be incorporated into the least-squares model to boost the generalization performance. Extensive experiments conducted on benchmark data collections verify the effectiveness of our proposed model.