D. Ruta (UK)
Classification, Classifier fusion, Visualisation
An explosion of many different classifier fusion techniques underlines the current state of the art in pattern recogni tion (PR). Rapid progress in this domain is being fuelled by the increasing commercial demand for the advanced data mining techniques capable of extracting useful knowl edge from terabytes of corporate data. Now, when com munication and processing power gradually becomes up for the job, more and more PR models is being reconsid ered for the profitable commercial applications. From the business perspective, classifier fusion techniques have been regarded as vastly complex and expensive black-box sys tems that sometimes perform well and sometimes not or even fail, over which the user has no control and under standing. Nevertheless research in this domain shows that multiple classifier systems provide the only option for fur ther performance improvement beyond the performance of the individual best model. This work intends to bridge the gap that prevents fusion systems from commercial applica tions and demonstrates a highly informative visual method that allows to understand and interpret how individual clas sification models are combined and how do they come to a decision of assigning patterns to different classes. The method allows to visualise individual discriminant func tions as well as their combinations using many different fusion techniques. For both, the user can observe the de cision boundaries among classes and see the landscape of the posterior class probability estimates in the whole input space. The method is tested on the illustrative examples of real benchmark dataset.
Important Links:
Go Back
