Experimental analysis and comparison of multilabel problem transformation methods for multimedia domain

Multi-label learning is the term used to express a type of supervised learning that requires classification algorithms to learn from a set of examples; each example can belong to one or multiple labels. The learning task consists of breaking the multi-label classification problem into several single label classification problems. This learning process results in the prediction of new class labels for a new example. Nowadays, the research community pays significant attention for Multi-label classification due to its relevance to many important domains including, video and audio, images and other media, text, and bioinformatics. Among the previously mentioned domains, Multimedia has the greatest part of interest in multi-label learning due to the increasing demand to efficiently access large collections of images and videos and developing applications that are used for indexing, searching and browsing multimedia data. In this paper, we present an analysis and experimental comparison of four multi-label learning methods applied to three multimedia benchmark datasets using five evaluation measures. In the experimental study, each method is applied to all datasets; alternatively, each problem transformation method is applied against all 54 classifiers in order to find the classifier that gives the best performance for each dataset and classification method.