Tuning model parameters in class-imbalanced learning with precision-recall curve.

Area of Science:

• Machine Learning
• Data Science
• Statistical Evaluation

Background:

• Class-imbalanced data presents challenges in model evaluation.
• Receiver Operating Characteristic (ROC) is commonly used, but may not be optimal for imbalanced datasets.
• Precision-Recall Curve (PRC) is less frequently employed despite its potential.

Purpose of the Study:

• To investigate the performance of PRC as an evaluation metric for class-imbalanced data.
• To compare the effectiveness of PRC against ROC in this context.
• To highlight the advantages of PRC for skewed datasets.

Main Methods:

• Comparative analysis of PRC and ROC metrics.
• Testing on a proposed algorithm with parameter tuning.
• Validation through simulation studies and real-world data examples.

Main Results:

• PRC demonstrates competitive performance with ROC for imbalanced data evaluation.
• PRC is effective in model parameter tuning for skewed datasets.
• PRC offers advantages over ROC in specific scenarios of imbalanced learning.

Conclusions:

• PRC is a competitive and effective alternative to ROC for class-imbalanced learning.
• PRC is suitable for preprocessing skewed data, including variable selection.
• PRC aids in building robust classifiers for imbalanced learning scenarios.