Research On Ensemble Learning With Multiple Views Based On Pairwise Constraints

In many practical problems data have some views which have complementarities informationand correlation information. We can improve the performance by mining that. Because of lack ofprior knowledge, it becomes very important to improve performance of learning method. As a goodmeans ensemble learning was added to it. Most existing methods train learners independently anditeratively from different views. The correlation information cannot be utilized in those methods.Canonical correlation analysis (CCA) is a method used to reveal linear relationships between twoviews. So we do a research on multi-view data based on CCA. In ensemble learning, diversityamong base classifiers is crucial. And most existing ensemble learning methods obtain diverseclassifiers by resampling instances or features, because the number of pairwise constraintsindicating whether paired instances belong to the same class or not, is much larger than that ofinstances, resampling pairwise constrains may get more diversity than resampling instances.Addressing the two problems above, we extend multi-view ensemble learning through pairwiseconstraints. At the same time, we also do a research in the situation of missing data in practicalapplications.In this thesis, we proposed a multi-view ensemble learning algorithm based pairwiseconstraints which is called ECAC. We give two versions of this algorithm. Firstly, we introduce theidea of constraints to multi-view ensemble learning. We randomly take constraints to get diversitywith two strategies. Then we make features extraction and fusion with CCA based constraints. Atlast we give a ensemble to the base classifiers. Accordingly, we can preserve correlations betweendifferent views and also combine the ensemble learning. The experiments carried out on multiplefeature database and facial databases showed the good performance of this algorithm.Secondly, we considered the situation of missing samples in this page. Because in traditionalCCA algorithms instances must appear in pair, the algorithm proposed above are not applicable.We make some change with these algorithms based correlations in and between classes. Then weget improved versions of these algorithms above. And the result of Experiments shows theeffectiveness of the proposed method.