Graph-constrained supervised dictionary learning for multi-label classification

In this work, we tackle the problem of multi-label classification using a sparsity-based approach. Multi-label classification problems, in which each instance is associated with a set of multiple labels, have received significant attention over the past few years due to the ongoing growth of data dimensions and availability. However, the dependency between labels poses new challenges to existing classification techniques. We propose a supervised dictionary learning algorithm suited for the multi-label setting. The suggested scheme introduces a novel graph Laplacian regularization that encapsulates the training set labels. This regularization explicitly takes into account the local manifold structure of the data, thus promoting the discriminative power of the learned sparse representations. Experiments on two different real-world multi-label learning problems, i.e. natural scene classification and yeast gene functional analysis, demonstrate that our proposed algorithm achieves superior performance to other dictionary based approaches as well as some established multi-label learning algorithms.