Research On Dynamic Ensemble Learning Algorithm Based On Attribute Generalization

Ensemble learning is one of research focuses in machine learning area in recent years.Its main idea is that, firstly use a part of base classifiers to classify, then collect theirclassification results and combine them by some way to obtain final result, which isgenerally more accurate than any of base classifiers. And, base classifiers are independentfrom each other. Dynamic ensemble technology is an important means which can furtherimprove the classification performance of ensemble system. There are two kinds ofdynamic ensemble technologies. One is to dynamically select part of base classifiers, not all,to participate in ensemble. And the other is to adjust the weights of base classifiers.Hierarchy characteristic of data is one of the nature characteristics. In this thesis, wemainly try to research how to use data’s hierarchy characteristic to construct dynamicensemble classifiers with high performance.This thesis provides some new insights into ensemble learning:(1) For the data set with hierarchy characteristic, an ensemble learning algorithm basedon generalized attribute value partitioning (GAVPEL) is proposed. Based on the hierarchystructure of the data, GAVPEL divides the training data set into smaller training data setswith multi-grain size in different levels using attribute generalization technology first, thentrain all generated training sets, one for each, to get a part of base classifiers. Whenclassifying a new instance, GAVPEL selects a part of base classifiers based on conditionattribute value and combines the results of these classifiers with majority voting.Experimental results show that GAVPEL is more efficient than traditional ensemblelearning algorithms such as Bagging and AdaBoost.(2) Since there are various generalization paths with different generalization ways, M2,improvement of GAVPEL is proposed. M2uses a part of generalization paths, whichintegrated in MRML model, to construct smaller training data sets, and similarly train baseclassifier on each smaller training data set. When classifying, M2selects all base classifierson multiple paths to participate in ensemble and the final result is generated using majorityvoting too. Experimental results show that M2can not only improve classificationperformance of ensemble system, but also improve its robustness.