Rough reducts based SVM ensemble

Neural network ensemble has demonstrated many advantages over single neural networks in terms of generalization ability and parameters configuration. As compared to traditional bagging and boosting that are typical of generating individual network by horizontally partitioning training dataset, this paper proposes a new ensemble method called RRSE (rough reducts based SVM ensemble) that differs in its individual-generating technique. RRSE generates individual SVM (support vector machine) of ensemble by projection of training dataset on sufficient and necessary attribute sets (reducts). For the sake of structured/semi-structured data we are faced with under most circumstances, RRSE employs rough set theory to preserve the conditional attributes' dependency on decision attributes and generate all minimal reducts. Because of the distinct significance and classification ability among reducts, each individual networks of RRSE ensemble accordingly learns with wider variance and less computation complexity, thus achieving good ensemble generalization ability and learning efficiency.