An empirical study of empty prediction of multi-label classification

• This is the first empirical study of empty prediction of multi-label classification.
• Every algorithm considered all made empty prediction on different datasets.
• HOMER and RAkEL have the overall lowest empty prediction rates in the study.
• Four thresholding methods which in theory can solve empty predictions are compared.
• Probabilistic thresholds are the best solution in terms of example based F1.

A detailed and extensive empirical study of empty prediction of multi-label classification is conducted in this paper and to the best of our knowledge this work is the first empirical study of this problem.Total 8 state of the art multi-label classification methods, BR, CC, CLR, HOMER, RAkEL, ECC, MLkNN, and BRkNN, are compared on 11 datasets. The empirical results clearly answer the two research questions, (1) whether empty prediction problems happen in commonly used state of the art multi-label classification methods and what their empty prediction rates (EPR) on different test sets are and (2) what multi-label classification methods are with overall highest/lowest EPRs. Specifically, it is empirically shown that every method considered all made empty predictions on different datasets. In addition, several thresholding methods which in theory can solve empty prediction are compared. The clear answers to the two research questions and the experimental findings are the main contributions of this work to multi-label classification.