In multi-label classification each instance can be associated to more than one label. For example, a music record can be associated to both labels 'happy' and 'relaxing'. Labels can be related with co-occurrence dependencies: for example, labels 'happy' and 'sad' can not be associated to the same music record. Labels can also be related with preference relations: for example, the label 'happy' is preferred over the label 'relaxing' to be associated to a music record containing several pikes. Label relations can help to better predict labels associated to instances. Existing approaches can learn either co-occurrence relations or preference relations. This work introduces an approach allowing to learn the two types of relations in order to improve the predictive performance. Experiments carried out show that the new introduced approach gives the best prediction results compared to five approaches from the state of the art.