Multi-label classification refers to the task that predicts one instance to be one or more labels in the set of labels. Nowadays, it is increasingly required by the real-world applications, such as text categorization, functional genomics and semantic scene classification. The main challenge for multilabel classification is the huge number of the labels. Traditional methods for multi-label classification include decomposing it into a set of independent binary labels or considering the pairwise relations between labels. Few works took the labels correlations into consideration. In this paper, a novel method for effectively exploiting the correlations between the labels is proposed. The Bayesian Network is used as the main tool to represent the correlations among the labels. It is constructed by the labels of each instance. For each label in the Bayesian network, the network is divided into two partitions: one is the nearest k-labels around the label; the other is the remaining labels. By this way, the task of learning the total set of labels is decomposed into the subtasks of learning a set of label powsets. A vote-based mechanism is introduced to predict the final proper set of labels for a new instance. Experiments on a board range of datasets validate the effectiveness of our BNNK method against the other well-established methods.
|Number of pages||10|
|Journal||Advances in Information Sciences and Service Sciences|
|Publication status||Published - 1 May 2012|
- Bayesian networks
- K-Nearest nodes
- Multi-Label classification