Banzhaf random forests: Cooperative game theory based random forests with consistency

Area of Science:

• Machine Learning
• Computational Statistics
• Game Theory

Background:

• Random forests are widely applied but lack theoretical depth.
• Existing methods like information gain ratio overlook feature interdependencies.

Purpose of the Study:

• To propose a novel random forests classification algorithm, Banzhaf random forests (BRFs).
• To enhance feature importance evaluation using cooperative game theory.
• To improve classification accuracy and theoretical understanding of random forests.

Main Methods:

• Developed a new random forests algorithm, BRFs, integrating cooperative game theory.
• Utilized the Banzhaf power index to assess feature importance by analyzing feature coalitions.
• Proved the consistency of the BRFs algorithm.

Main Results:

• BRFs demonstrate superior classification accuracy compared to existing consistent random forests.
• BRFs performance is comparable or better than Breiman's random forests, SVMs, and KNNs.
• The Banzhaf power index effectively captures feature dependencies.

Conclusions:

• BRFs offer a theoretically sound and practically effective advancement in random forests classification.
• The integration of cooperative game theory provides a novel approach to feature evaluation.
• BRFs successfully narrow the gap between random forests theory and real-world applications.