Probabilistic Predictions of Ensemble of Classifiers Combined with Dynamically Weighted Majority Vote

Eftim Zdravevski, Andrea Kulakov, Slobodan Kalajdziski, and Danco Davcev


Machine learning, prediction methods, pattern classification, decision-making


This paper presents a new method for dynamic calculation of weights that can be used in the process of aggregation of classifications by weighted majority vote. The proposed method can be used for all binary classification problems for classifiers that produce probabilistic classifications. Most aggregation functions produce an output which only represents the aggregated classification of an ensemble of classifiers and sometimes this isn't enough. This paper also proposes a method for estimation of the probability of an aggregated classification. The estimated probability of the aggregated classification is essential if the performance of the ensemble of classifiers needs to be expressed in terms of Area Under the Receiver Operating Curve or some other performance measures that classifications’ probability. The experimental results demonstrate the performance improvements obtained by applying the proposed methods to an ensemble of classifiers compared to individual classifiers.

Important Links:

Go Back