Multi-task Sparse Regression Metric Learning for Heterogeneous Classification

With the ubiquitous usage of digital devices, social networks and industrial sensors, heterogeneous data explosively increase. Metric learning can boost the classification performance via jointly learning a set of distance metrics from heterogeneous data. The metric learning algorithms are affected by the noisy doublets, i.e., the similar and dissimilar sample pairs. It is also a challenging issue to balance commonality and individuality for multi-view metric learning. To address the above issues, in this paper, we propose a novel multi-task group sparse regression metric learning (MT-SRML) for heterogeneous classification. Metric learning is formulated as sparse regression problem. The group sparse regularization on the repression coefficients of the doublets can restrain the effect of the noisy sample pairs jointly for multiple views. Experiments on heterogeneous data show that the proposed MT-SRML outperforms the state-of-the art metric learning algorithms in terms of both accuracy and efficiency.