A Hybrid Thresholding Strategy combining RCut and PCut for Multi-label Classification

Multi-label classification is a variant of the classification problem where multiple labels may be assigned to each instance. Usually multi-label classification algorithms output a numerical score for each label, indicative of their relevance to a query instance. However, in many applications the desired output is a bipartition of the labels into relevant and irrelevant w.r.t the query instance. Bipartitions can be obtained from scores using various thresholding strategies, such as PCut strategy which selects relevant instances per label, and RCut strategy which selects relevant labels per instance. However, we suggest that a combination of both strategies would provide better classification performance. In this paper, we propose a fuzzy-based approach to combine PCut and RCut strategies, by converting the crisp relevance into fuzzy one, merging them linearly, and defuzzifying again. Our experiments shows that our hybrid approach indeed outperforms both strategies.