Collaborative analysis in the combining of multiple classifiers .

摘要

Because of varying individual classifier performance, combinations of multiple classifiers are frequently used for pattern classification. This thesis introduces collaborative analysis as a means of analysing the results of multiple classifiers for the purpose of understanding the relationships between classes and classifiers. From this understanding, we can extract problem-specific knowledge that can increase classification accuracy. Collaborative analysis collects the interdependencies between class and classifier in data structures we have named Class-Classifier Relationship Forests. These derived relationships can give domain-specific insight. This work also shows that collaborative analysis of different classifiers is able to determine which of those classifiers, and in what combination, can increase the performance accuracy of classification. In comparing combinations of multiple classifiers through our method with the classifier combining method of majority voting, we show that collaborative analysis provided valuable insight into the relationship between the data and multiple classifiers, and increased classification accuracy.