Multi-label semi-supervised classification through optimum-path forest

Abstract Multi-label classification consists of assigning one or multiple classes to each sample in a given dataset. However, the project of a multi-label classifier is usually limited to a small number of supervised samples as compared to the number of all possible label combinations. This scenario favors semi-supervised learning methods, which can cope with the absence of supervised samples by adding unsupervised ones to the training set. Recently, we proposed a semi-supervised learning method based on optimum connectivity for single-label classification. In this work, we extend it for multi-label classification with considerable effectiveness gain. After a single-label data transformation, the method propagates labels from supervised to unsupervised samples, as in the original approach, by assuming that samples from the same class are more closely connected through sequences of nearby samples than samples from distinct classes. Given that the procedure is more reliable in high-density regions of the feature space, an additional step repropagates labels from the maxima of a probability density function to correct possible labeling errors from the previous step. Finally, the data transformation is reversed to obtain multiple labels per sample. The new approach is experimentally validated on several datasets in comparison with state-of-the-art methods.