Multi-label classification (MLC) generalises the conventional binary and multi-class classification by allowing instances to be linked with one or more of the class labels. Therefore, class labels in MLC are not mutual exclusive as in single label classification (SLC). Consequently, the search space of the MLC problem is large compared with that of SLC and grows in an exponential way. One main approach to solve MLC problem is through forcing instances to be associated with only one class label. This approach of handling the problem of MLC has been widely known as problem transformation method (PTM). Existing PTMs depend on the frequency of class labels as a transformation criterion, which causes several problems such as imbalance class distribution, complicating the training phase and most importantly decreasing the accuracy of the classification task. Therefore, in this paper, a new PTM is proposed based on the positive local dependencies among labels. The proposed PTM aims to facilitate capturing the most accurate positive dependencies among labels and hence improve the predictive performance of the classification task. Experiments on several datasets revealed the superiority of the proposed PTM compared with the existing PTMs, especially with high cardinality datasets.

No abstract

No abstract

No abstract