An enhancing framework with an emphasis on decision balance in ensemble regression

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Traditional ensemble regression learning (ERL) struggles with complex group decision-making (GDM) and instability due to diversity.
• These limitations have hindered theoretical advancements and practical applications in ERL.

Purpose of the Study:

• To propose a novel ERL framework with decision balance (DBERL) that integrates GDM and decision-balanced structures.
• To overcome the limitations of traditional ERL by enhancing stability and improving GDM processes.

Main Methods:

• DBERL models GDM as a decision-balanced network (DBN), adaptively clustering individuals for balanced structures.
• A hierarchical balanced attention mechanism (HBA) aggregates individual decision influences.
• A phased feedback mechanism promotes ensemble consensus.

Main Results:

• DBERL demonstrated superior performance across nine diverse datasets, outperforming 11 baseline models and 12 ensemble strategies.
• The framework achieved top ranks in statistical testing across six dimensions: fitting, correlation, interpretability, stability, data sensitivity, and generalization.
• Internal modules of DBERL were rigorously validated for effectiveness.

Conclusions:

• DBERL effectively addresses limitations in ERL by integrating GDM with decision-balanced structures.
• The proposed framework offers significant improvements in performance, stability, and generalization capabilities.
• DBERL represents a breakthrough in ERL, paving the way for further theoretical and applied research.