Does Classifier Fusion Improve the Overall Performance? Numerical Analysis of Data and Fusion Method Characteristics Influencing Classifier Fusion Performance

The reliability of complex or safety critical systems is of increasing importance in several application fields. In many cases, decisions evaluating situations or conditions are made. To ensure the high accuracy of these decisions, the assignments from different classifiers can be fused to one final decision to improve the decision performance in terms of given measures like accuracy or false alarm rate. Recent research results show that fusion methods not always outperform individual classifiers trained and optimized for a specific situation. Nevertheless fusion helps to ensure reliability and redundancy by combining the advantages of individual classifiers, even if some classifiers are not performing well for specific situations. Especially in unexpected (untrained) situations, fusion of more than one classifier allows to get a suitable decision, because of different behavior of classifiers in this case. Nevertheless, there are several examples, where fusion not always improves the overall accuracy of a decision. In this contribution fusion options are discussed to overcome the problem to overcome the aforementioned problem and to define influencing factors on overall fusion accuracy. As a results requirements for good or guaranteed or possibly increased fusion performance and also suggestions denoting those options not leading to any kind of improvement are given. For illustrating the effects a practical example based on three characteristics of fusion methods (type of classifier output, use of these outputs and necessity of training) and four data properties (number of classes, number of samples, entropy of classes and entropy of attributes) are considered and analyzed with 15 different benchmark data sets, which are classified with eight classification methods. The classification results are fused using seven fusion methods. From the discussion of the results it can be concluded, which fusion method performs best/worst for all data sets as well as which fusion method characteristic or data property has more or less positive/negative influence on the fusion performance in comparison to the best base classifier.Using this information, suitable fusion methods can be selected or data sets can be adapted to improve the reliability of decisions made in complex or safety critical systems.