Exploitation of label relationship in multi-label learning
Zhi‐Hua Zhou · 2012
Summary form only given. Traditional supervised learning deals with problems where one instance is associated with a single class label, whereas in many real tasks, one instance may be associated with multiple class labels simultaneously. The paper introduces some of our recent findings in the exploitation of label relationship. First, it introduces an approach which is able to discover and exploit label relationship automatically. This approach does not require label relationship as input for the model construction process; instead, it is able to provide not only accurate predictions, but also reasonable estimate of label relationship. The label relationship discovered is usually asymmetric, which is quite different from symmetric label relationship assumed by previous studies, but is more consistent with realistic situations. Secondly, it advocates to exploit label correlations in the data locally, because in real tasks it is usually the case that the label correlation is shared by only a subset of instances rather than all instances.