H. Prehn and G. Sommer (Germany)
Incremental Classification, Local Credibility Criterion
In this paper we propose the Local Credibility Concept (LCC), a novel technique for incremental classifiers. It measures the classification rate of the classifier’s local models and ensures that the models do not cross the borders between classes, but allows them to develop freely within the domain of their own class. Thus, we reduce the depen dency on the order of training samples, an inherent prob lem of incremental methods, and make the classifier robust w.r.t. selecting the algorithm’s parameters. These only in fluence the number of models, whereas the performance is controlled by the LCC automatically on a local scale. In contrast to other algorithms, the models of our method are more adaptable as they can also shrink and vanish. This al lows classes to move their domains in the data space mak ing the LCC-Classifier also applicable to drifting data con cepts. We present experiments to demonstrate these capa bilities as well as some benchmark tests that show the al gorithm’s competitive performance.
Important Links:
Go Back
