Improving the Performance of Multi-Label Classifiers via Label Space Reduction

Multi-label classification is related to the problem of learning a predictive model from examples that may be associated with a set of labels simultaneously. The learning process in datasets with large label spaces turns into a really challenging task since the computational complexity of most algorithms depends on the number of existing labels. This paper proposes a methodology for reducing the label space a predefined percentage of labels, with the aim of improving the runtime of the multi-label algorithms without producing a significant variation in the predictive performance. The experimental analysis demonstrates a drastic reduction in runtime, while proving that in many cases, the reduction of the label space up to 50% does not significantly affect the performance using four well-known evaluation measures.