Análise das correlações entre as métricas de avaliação de desempenho de classificadores multirrótulo

Abstract

In the field of machine learning, multi-label classification is a variant of the classification problem in which multiple labels may be associated with each instance. Multi-label classification requires classification algorithms which need the use of measures to evaluate their performance. In the literature of machine learning, it was not found studies which point to the existence of correlations between performance measures for multi-label classifiers. Such studies are important because they can assist researchers in the field in order to support decision making on which algorithms may be chosen or considered for certain problem transformation approaches. In this context, this thesis presents a study of performance measures for multi-label classification algorithms and their correlations. The main goal of this research was to identify correlations between the performance measures for multi-label classifiers. In order to reach the main goal, it was necessary to use: ten multi-label datasets; five problem transformation approaches (BR, CC, LP, PS, and RAkEL); five base classifiers (J48, KNN, NB, SVM, and RIPPER); and, twelve performance measures (HLoss, SAcc, Prec, Rec, FM, Acc, AvPrec, Cov, 1-Err, IsErr, ErrSS, e RLoss). The machine learning validation technique used for experiments was the 10-fold cross-validation. In order to run experiments and calculate performance measures, the JAVA programming language and the MULAN library was used. Correlations were calculated by both Pearson's and Spearman's correlation coefficients. After analyzing the results, it was possible to conclude that there are as linear as non-linear correlations between the selected performance measures. Thus, all correlations found made it possible to identify classification algorithms which are more suitable to be used together with the problem transformation approaches in experiments related to multi-label classification problems