Evidence Combination for Baseline Accuracy Determination

Several classifier combination approaches have been proposed in machine learning literature in order to enhance the performance of simple learning schemes. This paper presents a new classifier fusion system based on the principles of the Dempster-Shafer theory of evidence combination. The system tackles the advantages of combining different sources of information to attain a high degree of stability across different problem domains. The uncertainty evaluation provided by the Dempster-Shafer theory also contributes to achieving this stability. System evaluation has confirmed the assumptions related to stability and allows us to formulate a method of establishing the baseline accuracy for any problem domain. Thus, the choice of a specific learning scheme for a certain problem is justified only if it's performance is better than that of the system proposed here.