A Semi-Supervised Ensemble Approach for Multi-label Learning

In this paper, we present a new ensemble approach for semi-supervised multi-label classification which exploits both the dependencies between the class labels and the unlabeled instances to enhance the multi-label classification performance. Our approach combines both data resampling (bagging) and label random subspace strategies for generating a committee of multi-label models in a co-training style algorithm. The key ideas behind this approach are to i) promote and maintain diversity in the multi-label base-classifiers committee, ii) define a new cost oriented metric to estimate the prediction confidence for each label, and iii) use a new multi-label out-of-bag feature importance measure that makes full use of labeled and unlabeled in the semi-supervised setting. Experimental results on various benchmark data sets approved that the proposed approach outperforms recent state-of-the-art supervised and semi-supervised multi-label algorithms over different multi-label metrics.