Knowledge-based methods for classifier combination: an experimental investigation

Many combination methods have been proposed so far for classifier combination. In order to achieve better performance, some methods also use a-priori knowledge on the set of classifiers. Unfortunately, in this case the effectiveness of the methods is very difficult to predict since there is little assurance that the results obtained in controlled tests can be obtained under different working conditions imposed by the real applications. In this paper, the role of a-priori knowledge in classifier combination is evaluated. A recent methodology is used for the analysis of methods for classifier combination. The performance of a combination method is measured under different working conditions by simulating sets of classifiers with different characteristics for the test. A random variable is used to simulate each classifier while a suitable estimator of stochastic correlation is used to measure the agreement among classifiers.