Selection Of Base Classifiers In Combination Of Multiple Classifiers

In the field of data mining, classification has always been a most important technology. However, many classification techniques that have been introduced are of single classifier method. Because the improvement of single classifier's accuracy has reached the bottleneck, so people propose the concept of the combination of multiple classifiers, which means using multiple classifies to classify the instance and combine every classifier's result to give a final result. Experiments show that combination can improve the classifier's performance obviously. Therefore, couducting research on classifiers combination is of significant theoretical value and practical benefit.Firstly, this paper gives a comprehensive description of every research direction relating to classifiers combination, including basic theory and representative methods. The issues included are what is combination and some critical questions, such as the diversity of combination, the strategy of creating base classifier, how to select base classifiers and how to combine the choosed base classifiers.Based on the existed reseach, this paper proposes two particular algorithms about selection of base classifiers. Firstly, how to use diveristy to create combination becomes a research focus. A new selection method using diversity is proposed firstly based on the Boosting method. Furthermore, this paper analyzes the boosting method and point out its potential disadvantage, and proposes another new method that can dynamically allocate weights to base classifiers using current test instance, in order to overcome the defect of static weight allocation in boosting method. Finally, this paper describes weka, the platform used in experiment. Then it implements the algorithms mentioned above and gives a detail on the original implementment of meta-learning in weka. After that, this paper gives the experiment results of the algorighms and some classical methods and compares the difference between those results. The results show that the algorithms proposed in this paper have better performance in majority of the dataset.