Simultaneous Nonlinear Label-Instance Embedding for Multi-label Classification

In this paper, unlike previous many linear embedding methods, we propose a non-linear embedding method for multi-label classification. The algorithm embeds both instances and labels into the same space, reflecting label-instance relationship, label-label relationship and instance-instance relationship as faithfully as possible, simultaneously. Such an embedding into two-dimensional space is useful for simultaneous visualization of instances and labels. In addition linear and nonlinear mapping methods of a testing instance are also proposed for multi-label classification. The experiments on thirteen benchmark datasets showed that the proposed algorithm can deal with better small-scale problems, especially in the number of instances, compared with the state-of-the-art algorithms.