Joint-optimized feature selection and classifier hyperparameters by salp swarm algorithm in piano score difficulty measurement problem

This study proposes the semantic-explicit features that characterize difficulty, and jointly optimizes feature selection and classifier hyperparameters by the salp swarm algorithm (SSA) to deal with the corresponding mixed-integer programming problem with constructing large-scale piano score difficulty level datasets. The difficulty level of piano scores is essential for piano learners to choose the appropriate piece, especially for beginners and amateurs. However, the previous studies lack an open-access baseline dataset and sufficient difficulty-related features, as well as the separate optimization of and feature and model hyperparameter. To address such problems, this study constructs large-scale difficulty-level datasets, proposes novel difficulty-related features, and jointly optimizes feature selection and classifier hyperparameters due to the coupled effect of feature selection and model optimization. The search space of the joint optimization is complex due to the strong mutual constraint relationship of difficulty levels in the piano-score difficulty measurement (PSDM) problem. SSA is adapted to the joint optimization scheme of the PSDM problem with the advantages of only one main controlling parameter and less computation complexity involving the gradual SSA movement approach to balance global exploration and local exploitation in an unknown and complex search space. The joint-optimization mechanism by SSA achieves an overall accuracy of 78.80% and 60.68% on two datasets of 677 and 2040 piano pieces with difficulty levels of four and nine, respectively. The results of recognition accuracy obviously validate the distinguished performance of our joint-optimization scheme compared to the successive optimization and joint optimization by other seven optimization algorithms in terms of the PSDM problem.