On the Combination of Two Decompositive Multi-Label Classification Methods

Abstract

In this paper, we compare and combine two approaches formulti-label classi cation that both decompose the initial problem intosets of smaller problems. The Calibrated Label Ranking approach isbased on interpreting the multi-label problem as a preference learningproblem and decomposes it into a quadratic number of binary classi ers.The HOMER approach reduces the original problem into a hierarchy ofconsiderably simpler multi-label problems. Experimental results indicatethat the use of HOMER is bene cial for the pairwise preference-basedapproach in terms of computational cost and quality of prediction.