Learning hidden information: SVM+

In this talk, I consider a new setting of the learning problems, where on the training stage one can use information that will be hidden on the test stage. It requires one to construct methods of learning which using both hidden and non-hidden information on the training stage, will be able to construct a rule that on the test stage uses only non-hidden information, and performs better than rules that can be obtained by conventional learning methods. I will introduce a new learning methods (both for pattern recognition and regression estimation) called SVM+ that generalises SVM and can solve this problem. The detailed description of SVM+ one can find in the Afterword of the 2nd Edition of my book Estimation of Dependencies Based on Empirical Data, Springer 2006.