A Study On The Multiple Classifier Systems

Because multiple classifier systems (MCSs) might reach much better performance and robustness than a single optimum classifier, the technology of MCSs fusion has received more and more attention and made greater progress in recent years. However, it also brings a big challenge to researchers working in this area. Much more researchers try to solve questions such as how to use each classifier's experitise effectively and how to obtain support from middle concepts to seek regional superiority and so on. Aiming to improve the performance of MCSs, this thesis starts with a study on MCSs - with a brief introduction on existing algorithms together with a comparison of different combination strategies. Two novel classifiers fusion methods are peoposed and developed, which are different from each other in classifiers combination and selection. The first one makes use of class-wise expertise of each classifier, and generates a combined classification system. The experimental results evaluated on some toxicity datasets show that the proposed method outperforms each member classifier, both in effectiveness and robustness. Another approach works on a transformed data space of multiple classifier behavior vectors instead of the original data space. A tentative fusion algorithm named DRC is proposed which transforms the original space into a new space via the classifier behavior, followed with a reduction and learning model built in the new space. It is proved by the experiments that the performance of DRC is better than that of each classifier both in the original space and in the new transformed space.The assumption of independence among classifiers has been a key issue of classifier combination research. An approach choosing classifier member based on class-wise expertise and information gain is presented, which aims to improve the performance of the combined system. Evalution on some real-world datasets shows that the proposed classifier selection strategy gains a good result.