RB: A New Method for Constructing Multi-Label Classifiers Based on Random Selection and Bagging

Abstract

In many real world prediction problems, a classifier must, or should, assign more than one label to an instance, e.g. prediction of machine failures, musical genre classification, etc. For this kind of problem, multi-label classification methods are needed. One approach frequently used to learn multi-label predictors divides the problem into one or more multi-class classification problems, and combines the models constructed for each sub-problem to classify new instances with multiple labels. Although there are many multi-label learning methods, there is a need for exploring methods that can lead to improvement in prediction power. In this work, we propose and evaluate a new method, called RB (Random-Bagging), based on dataset transformation and combination of classifiers. Six real-world datasets were used to evaluate our method, which was compared to three existing methods. Results were considered promising.

Cardinality and Density Measures and Their Influence to Multi-Label Learning Methods

“…Correlation between the results and Card, as well as between the results and Dens, was expected. However, as we explain later in this section, we observed that many situations where correlations were observed in [5] could not be observed in our results. We also calculated Spearman's rank correlation between Card and each measure results, and also was calculated between Dens and each measure results using (i) only natural datasets; and (ii) only MSD-based datasets.…”

“…We present the results obtained for MSD-based datasets. We also bring to this work the results obtained for the six datasets used in [5] to enrich our analysis, and we also enlarge that results to evaluate similarly to the evaluation using MSD. We analyze the relation between (i) cardinality and (ii) density and the results obtained by each method.…”

“…In this work, we use the Million Song Dataset. Also, we use and expand the results obtained in [5], which consider other multi-label dataset found on the internet. We describe all these datasets in what follows.…”

“…Timbre is usually defined as the color of the sound, because by timbre we can identify a sound produced by different fonts, such as two musical instruments playing the same accord or two people singing the same melody [23]. Other important features are artist name (the singer of the music), title of the music, location (where the music was recorded), year when 53-71, 2014 c Brazilian Society on Computational Intelligence the music was recorded, time duration, segments-start, bars start, similar artists, terms and mbtags -MusicBrainz tags, provided by MusicBrainz 5 . The last five listed features, jointly to segments-timbres and segments-pitches, are multi-valued.…”

“…2. It is worth to mention that we extended the experiments described in [5] to appropriately analyze the impact of cardinality and density measures to both natural and MSD-based datasets. Figures 4 and 5 show, respectively, cardinality and density values of each dataset used in this work.…”