Research And Application Of Ensemble Learning Based On Conditional Mutual Information

Ensemble learning is the first of the four major research directions. Along with involving in researching it by more and more researchers, the two main algorithms Bagging and Boosting have been studied in a deep way. So some other methods should be found to devote to the development of ensemble learning. Selective ensemble learning is a special paradigm of ensemble learning. It ensembles distinct and strongest generalization individual learners with optimized method, and gains the better performance than that of integrating all individual learners. Currently, ensemble learning has been used in many fields such as planetary exploration, seismic analysis, Web information filtering, biometric identification, computer-aided medical diagnosis, etc.In this paper, a detailed analysis of theory and algorithms about ensemble learning is given, and then it probes into the reason why ensemble learning works and its advantages which are superior to traditional methods. After that, it points out the shortcomings and development directions of ensemble learning. At last, an introduction to the theory and main algorithms of selective ensemble learning is followed. Now there are the main two works that are finished in this paper as follows.Firstly, it introduces the theory of conditional mutual information in information theory, and then combines it with ensemble learning to construct a novel method called CMISEN (Conditional Mutual Information Based Selective Ensemble) to select the optimized individual learners. This method considers the influence among individual learners, and can prevent selecting the redundant individual learner to make the differences among individual learners that have been selected much bigger. This novel method is implemented in weak environment in Java language. In order to compare this method with the existing methods such as Bagging and MISEN (Mutual Information Based Selective Ensemble) in performance, we use the UCI standard data set to verify their performance. The result proves that the novel method is superior to the existing method Bagging and MISEN.Secondly, sensory evaluation is the hot topic in current days. In the field of tobacco, many computer intelligent methods have been used to solve problems. Currently, single models have been used to solve the problems, but the single model is not stable, it can gain a good performance for these data sets and bad for those data sets. If this single model is over-fitting on a data set, the generalization of this single model is low. So ensemble learning descends the risk that over-fitting makes the generalization of a single model become low. But the traditional single individual learner is not enough to improve the performance. So in this paper, it uses ensemble learning to implement sensory evaluation of tobacco. The result of experiment shows the ensemble learning is effective and superior to the former methods.At last, it gives the prospect of next work and views of the development of ensemble learning.