Research Of Ensemble Selection Algorithm Based On Semi-Supervised Regression And Its Application

As a new paradigm of machine learning, Ensemble learning uses multiple learners to solve the same problem, which could significantly improve the generalization ability of learning systems. Due to its good performance and broad applicability, Ensemble learning has become a hot research topic in the machine learning community. It has been widely used in different fields, such as planet exploration, seismic wave analysis, web information filtering, biology feature recognition, and computer aided medical diagnoses. However, Traditional researches of ensemble learning are based on the supervised learning which needs a large number of labeled training examples. In practice, to obtain labeled training examples is time-consuming and costly during the training process. And Ensemble learning can not get good performance with only a few labeled training examples. Therefore it's worth studying to improve the performance of ensemble learning with a few labeled training examples.Against to there are a lot of unlabeled examples, we combine ensemble selection with semi-supervised learning to propose a new ensemble selection algorithm based on semi-supervised regression namely SSRES. On one hand, the method uses a large number of unlabeled examples to reduce the requirement of labeled examples; on the other hand, it selects the learners by GRES to improve the generalization ability of learners. In this thesis, we mainly study the problem of ensemble selection based on semi-supervised regression. The main worksof this thesis are summarized as follows:First, researching on current existing ensemble learning, we prove the good performances of Boosting and Bagging on the Weka database. It also shows that ensemble learning can improve the performance of learners. Second, we research on the theory of ensemble selection and semi-supervised learning. Then we implement GRES and COREG on the development platform of Eclipse + Weka. We verify its validity on the Weka database and the synthetic database.Third, upon the above studying, we propose a new ensemble selection algorithm based on semi-supervised regression (SSRES) in order to solve the problem with a few of labeled examples. Then we implement the algorithm on the development platform of Eclipse + Weka, and analyze its performance on the Weka database and the synthetic database. The experiment results show it can improve the performance of learners by using unlabeled example and ensemble selection.In the end, against the feature of there may be a large number of untested and just a little of tested mix proportion data in concrete industry, we apply the algorithm of SSRES to strength prediction of concrete in order to improve its accuracy. The experiment results based on the real mix proportion data show that the proposed algorithm can improve the prediction accuracy, thus it can contribute to the achievement of the mix optimization.