Bagging-like metric learning for support vector regression

Metric plays an important role in machine learning and pattern recognition. Though many available off-the-shelf metrics can be selected to achieve some learning tasks at hand such as for k-nearest neighbor classification and k-means clustering, such a selection is not necessarily always appropriate due to its independence on data itself. It has been proved that a task-dependent metric learned from the given data can yield more beneficial learning performance. Inspired by such success, we focus on learning an embedded metric specially for support vector regression and present a corresponding learning algorithm termed as SVRML, which both minimizes the error on the validation dataset and simultaneously enforces the sparsity on the learned metric matrix. Further taking the learned metric (positive semi-definite matrix) as a base learner, we develop a bagging-like effective ensemble metric learning framework in which the resampling mechanism of original bagging is specially modified for SVRML. Experiments on various datasets demonstrate that our method outperforms the single and bagging-based ensemble metric learnings for support vector regression.