A novel framework based on the multi-label classification for dynamic selection of classifiers.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multi-classifier systems (MCSs) enhance predictive modeling by combining multiple classifiers.
• Dynamic selection (DS) techniques aim to improve MCS performance by selecting the most competent classifiers for each test sample.
• Traditional DS methods rely on detecting local data regions and measuring classifier competence, which can be complex.

Purpose of the Study:

• To propose a novel dynamic selection technique for MCSs that simplifies the classifier selection process.
• To introduce the first framework utilizing multi-label classification for dynamic classifier selection.
• To offer a computationally efficient and easy-to-implement alternative to existing DS methods.

Main Methods:

• A multi-label classifier is trained to directly determine the optimal set of classifiers for a given task.
• The proposed method bypasses the need for explicit competence measures or local region detection.
• In the generalization phase, the trained multi-label classifier predicts the appropriate subset of classifiers for test samples.

Main Results:

• The proposed multi-label-based dynamic selection technique achieves high classification accuracy.
• The method demonstrates significant simplicity in implementation and requires no meta-features.
• Statistical tests confirm the proposed method's superior performance compared to benchmark DS techniques.

Conclusions:

• The novel multi-label classification approach offers an effective and simplified solution for dynamic classifier selection in MCSs.
• This framework represents a significant advancement in meta-learning for ensemble selection.
• The technique provides a strong balance between performance and ease of use, outperforming existing methods.