Multi-label Classification Using Hypergraph Orthonormalized Partial Least Squares

In many real-world applications, human-generated data like images are often associated with several semantic topics simultaneously, called multi-label data, which poses a great challenge for classification in such scenarios. Since the topics are always not independent, it is very useful to respect the correlations among different topics for performing better classification on multi-label data. Hence, in this paper, we propose a novel method named Hypergraph Orthonormalized Partial Least Squares (HOPLS) for multi-label classification. It is fundamentally based on partial least squares with orthogonal constraints. Our approach takes into account the high-order relations among multiple labels through constructing a hypergraph, thus providing more discriminant information for training a promising multi-label classification model. Specifically, we consider such complex label relations via enforcing a regularization term on the objective function to control the model complexity and balance its contribution. Furthermore, we show that the optimal solution can be readily derived from solving a generalized eigenvalue problem. Experiments were carried out on several multi-label data sets to demonstrate the superiority of the proposed method.