A Correction Method of a Binary Classifier Applied to Multi-Label Pairwise Models

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multi-label (ML) classification presents challenges, especially with the label-pairwise (LPW) scheme.
• Existing LPW ensembles rely on binary classifiers that can be improved.

Purpose of the Study:

• To propose a novel correction method for binary classifiers within LPW ensembles.
• To enhance the performance of ML classification by addressing classifier inaccuracies.

Main Methods:

• A probabilistic (randomized) model is used to correct binary classifiers.
• The correction leverages a randomized reference classifier (RRC) and a local soft confusion matrix.
• Special cases for imbalanced labels and double-labeled instances are considered.

Main Results:

• The proposed correction methods were evaluated on 29 benchmark datasets.
• Experiments showed significant improvements in zero-one loss and Hamming loss compared to original LPW classifiers.
• The study investigated the relationship between classification quality and dataset characteristics like label imbalance and density.

Conclusions:

• The developed correction scheme offers a significant advancement for LPW-based multi-label classification.
• The method demonstrates superior performance, especially in reducing specific error metrics.
• Further research can explore the impact of dataset characteristics on the proposed correction approach.