Many real-world classifications associate simultaneously each data object to multiple labels. This kind of classification, called multi-label learning, is relevant to many domains, such as text categorization, gene function analysis, image annotation, drug discovery. Researchers have already proposed a variety of approaches to deal with multi-label data. However, data sources from the real world are growing ever larger and the multi-label task is particularly sensitive to the associated problems such datasets: complexity of correlations among larger sets of labels, high imbalance level in the labels and the large amount of unlabeled instances. This thesis deals with methods which exhibit high predictive performance in such multi-labels datasets.