In multi-label datasets, the number of labels associated with each instance is an important feature to be observed. Two relevant characteristics related to datasets' number of labels are cardinality and density. In this work, we use artificial datasets generated through a framework named Mldatagem, freely-available in the internet. This framework enables configuring some other characteristics of the generated datasets. In this paper we present a study that analyze how and when distinct characteristics of the datasets influence the performance of multi-label learning methods.