Collective decision efficiency and optimal voting mechanisms: A comprehensive overview for multi-classifier models

TL;DR

This paper introduces a game-theoretic framework for designing optimal voting mechanisms in multi-classifier models, enhancing collective decision accuracy and efficiency through weighted majority rules and adaptive strategies.

Contribution

It develops a comprehensive game-theoretic approach to voting in classifier ensembles, proposing new optimal rules and adaptive methods for improved collective decision-making.

Findings

Weighted majority rules are theoretically optimal for coalition-based voting.

A generalized risk-based framework enables classifier combination in pattern recognition.

Adaptive WMR adjusts to non-symmetrical feature space properties.

Abstract

A new game-theoretic approach for combining multiple classifiers is proposed. A short introduction in Game Theory and coalitions illustrate the way any collective decision scheme can be viewed as a competitive game of coalitions that are formed naturally when players state their preferences. The winning conditions and the voting power of each player are studied under the scope of voting power indices, as well and the collective competence of the group. Coalitions and power indices are presented in relation to the Condorcet criterion of optimality in voting systems, and weighted Borda count models are asserted as a way to implement them in practice. A special case of coalition games, the weighted majority games (WMG) are presented as a restricted realization in dichotomy choice situations. As a result, the weighted majority rules (WMR), an extended version of the simple majority rules, are asserted as the theoretically optimal and complete solution to this type of coalition gaming. Subsequently, a generalized version of WMRs is suggested as the means to design a voting system that is optimal in the sense of both the correct classification criterion and the Condorcet efficiency criterion. In the scope of Pattern Recognition, a generalized risk-based approach is proposed as the framework upon which any classifier combination scheme can be applied. A new fully adaptive version of WMRs is proposed as a statistically invariant way of adjusting the design process of the optimal WMR to the arbitrary non-symmetrical properties of the underlying feature space. SVM theory is associated with properties and conclusions that emerge from the game-theoretic approach of the classification in general, while the theoretical and practical implications of employing SVM experts in WMR combination schemes are briefly discussed. Finally, a summary of the most important issues for further research is presented.