An Improvement of Label Power Set Method Based on Priority Label Transformation

The automatic text categorization and the medical diagnosis were the first domain of applications that requires Multi-label classification. It consists to assign more than one label for each object. Later, its applications were widely increased to cover additional fields like functional genomics, music, biology, scene, video etc... There are two main categories of approaches for Multi-label classification: adaptation and transformation. We are interested in this paper in the transformation. Three most important techniques are considered for transformation (Binary Relevance, Label Power-Set and Classifier Chain). They transform multi-label into one or more appropriate single-label classification. We point out on three important limitations for the transformation to handle them in our contribution: a) the dependency between labels, b) the complexity of the algorithms, and c) the choice of single-label classifier. In this paper, our contribution consists of introducing a new transformation method called Label Priority Power-set (LPP). It is an improve version of Label Power Set that solves the three limitations. It transforms multi-label to multi-class learning. LPP solves the problem of label dependency by ordering labels according to their importance. Furthermore, LPP can be applied for dataset with a big number of labels. It is an efficient and accurate method that increases the performance and reduces the complexity. It applies decision tree as single label classifier in order to have the best choice of traditional classifiers. Experiment shows that our new method gave better results than the other techniques in terms of some of the evaluation metrics and complexity.