Multi-Label Classification: A Novel approach using decision trees for learning Label-relations and preventing cyclical dependencies: Relations Recognition and Removing Cycles (3RC)

Abstract

Multi-Label Classification (MLC) is a field of machine learning, which consists of classifying data by assigning to each instance a set of labels instead of one. These labels or classes can have dependencies between them. Omit this information can affect the predictive quality of classification. Considering these dependencies or ignoring them, when building the classifier, each has its drawbacks. The first approach facilitates the spread of learning errors and increases complexity of the task, especially if there are cyclical relationships between classes. While the second approach can give inconsistent predictions. There are multiple approaches designed to solve multi-label classification tasks, some of them take into consideration labels dependencies and others consider them independent. A new approach called PSI-MC proposes a novel way to learn the relations between labels without fixing a predefined structure. We propose an approach that uses the same principle as the PSI- MC, and which improves the way to eliminate cycles. Finally, we present the results of testing our new approach on four different datasets. According to four measures, our proposed approach called (3RC) is much better than binary relevance, RAKEL and MLKNN approaches.