ABSTRACT
It is usually assumed that each instance has only one label. Let X denote the instance space and L be a set of labels, a single-label training set could then be denoted as D = {(x1, l1), (x2, l2), . . . , (xn, ln)}, where xi is the ith instance and li is its relevant label taken from L. The objective of classifi cation could be viewed as learning a mapping from the instance space to the label space: f : X ‚ Y, based on a training dataset D. Generally, this kind of learning is also called single-label classifi cation [1].
