Semi-Supervised Multi-Label Classification Through Topological Active Learning

Abstract

Multi-label classification is becoming increasingly widespread as a data mining technique. Its objective is to categorize models in several non-exclusive groups, and is applied in such areas as news categorization, image labeling and music classification, among others. Our contribution is to use the paradigm of active learning with the topological power of the Act-SOM for semi-supervised multi-label classification, taking into account the multi-label information, and selecting unlabeled data which can lead to the largest reduction of the expected model loss. This paper deals with various multi-label datasets by presenting in active learning, a set of results ranging from: 1) Transductive classifier TSVM with relevance sampling methods for multi-labeled data in various application domains; 2) Proposed semi-supervised classifier Act-SOM in multi-label active learning, adopting a strategy relative to the evaluation by the uncertainty of the labels. Act-SOM based on active learning selects the most uncertain data while clearly improving the test rate with less than 30% of labeled instances added, which is our main contribution. We present the results from the statistical tests using critical diagrams. Thus, potential of the proposed multi-label classification method is demonstrate, due mainly to the competitive properties with global consistency of the semi-supervised Act-SOM through topological active learning.