Using informational confidence values for classifier combination: an experiment with combined on-line/off-line Japanese character recognition

Classifier combination has turned out to be a powerful tool for achieving high recognition rates, especially in fields where the development of a powerful single classifier system requires considerable efforts. However, the intensive investigation of multiple classifier systems has not resulted in a convincing theoretical foundation yet. Lacking proper mathematical concepts, many systems still use empirical heuristics and ad hoc combination schemes. The paper presents an information-theoretical framework for combining confidence values generated by different classifiers. The main idea is to normalize each confidence value in such a way that it equals its informational content. Based on Shannon's notion of information, the author measure information by means of a performance function that estimates the classification performance for each confidence value on an evaluation set. Having equalized each confidence value with the information actually conveyed, the author can use the elementary sum-rule to combine confidence values of different classifiers. Experiments for combined on-line/off-line Japanese character recognition show clear improvements over the best single recognition rate.