Paired Comparisons Method for Solving Multi-Label Learning Problem

Abstract

Multi-label classification problem is a further generalization of traditional multi-class learning problem. In multi-label case the classes are not mutually exclusive and any sample may belong to several classes at the same time. Such problems occur in many important applications (in bioinformatics, text categorization, intrusion detection, etc.). In this paper we propose a new method for solving multi-label learning problem, based on paired comparisons approach. In this method each pair of possibly overlapping classes is separated by two probabilistic binary classifiers, which isolate the overlapping and non-overlapping areas. Then individual probabilities generated by binary classifiers are combined together to estimate final class probabilities fitting extended Bradley-Terry model with ties. Experimental performance evaluation on well-known multi-label benchmark datasets has demonstrated the outstanding accuracy results of the proposed method.

Multi-label learning with incomplete class assignments

“…First, we observe that LIBSVM+Platt significantly improves the performance of LIBSVM in all four settings. This is consistent with [25], where the conversion procedure makes the outputs from different SVM classifiers more comparable and consequently leads to better performance for multilabel ranking. On the other hand, both LIBSVM and LIBSVM+Platt are outperformed by the other two multi-label learning methods, indicating the importance of developing multi-label ranking methods for multi-label learning.…”

Classifier Chains for Multi-label Classification

“…This is the one-vs-one approach, as opposed to the one-vs-rest approach used by BM, therefore requiring |L| 2 classifiers as opposed to |L|. [7] accompanies each pairwise classifier with two probabilistic models to isolate the overlapping feature space. They cite a computational bottleneck for this method for large datasets.…”

Classifier chains for multi-label classification

“…Although they can work with textual datasets, they scale poorly with the number of labels and can fail to perform well on sparse data (Freund and Schapire 1999). This family of methods has been recognised as having high computational complexity (Petrovskiy 2006).…”

“…CLR ) is a well-known example (many other pairwise methods are specific to the label ranking problem, which is beyond the scope of this paper). In Petrovskiy (2006), each pairwise classifier is accompanied by two probabilistic models to isolate the overlapping attribute space. This adds further complexity, and causes a computational bottleneck on large datasets.…”