Dynamic classifier combination method based on minimum cost criterion

This paper presents a novel method for multi-classifier dynamic corrbinati on based on minimum cost criterion. Different from common combination methods, the proposed dynamic combination selects the most suitable group of classifiers according to the Performance Predication Feature (PPF) extracted from the input sample. PPFs are the features of sample that have great influence on the performance of class fiers being studied. The decision is made based on the criterion that the selected group of classifiers should minimize the cos caused by recognition error and recognition on time. Systematic methods for making this kind of combination is proposed and a practical example of application is given. Because the adjustment of cost function will result in di fferent trade-off between recognition rate and recognition speed, it is very convenient to satisfy di fferent needs. The application in on-line Chinese character recognition technology shows that this kind of combination method has the merits of high flexibility and practicality, and it is indeed able to improve the system performance.