New Presence-Dependent Binary Similarity Measures for Pairwise Label Comparisons in Multi-label Classification

Similarity measures play an important role in data analysis, as the performance of different classification or clustering techniques relies on choosing an appropriate measure. Most formulas either increase values with “negative matches” or depreciate values by “absence mismatches”. As a result, to verify label dependencies in many data mining problems, e.g., multi-label classification, researchers usually must accept a compromise solution. In the paper, two effective formulas for binary vector comparisons are proposed. Both coefficients exclude “negative matches” but include “absence mismatches”. The formulas’ performance was examined and compared to other well-known binary similarity measures, including Jaccard and Dice or the Chi-square test for proportions. The experiments were performed on eight datasets of different characteristics regarding the multi-label classification problem and label dependencies. The investigations presented good efficiency of both new metrics in finding similarities between binary vectors and outperformed other examined metrics.