Research Of Reduced Multiple Kernel Support Vector Machine

In recent years, Support vector machine (SVM) based on a single kernel is not suitable for complex and large scale data due to the increasing complexity of the machine learning problems. In order to deal with large-scale complex heterogeneous datasets and improve the generalization ability of SVM, Multiple Kernel Learning (MKL, also called Multiple Kernel Support Vector Machine) based on multiple kernels has been proposed several years ago. However, the time complexity of training MKL increases greatly due to the introduction of multiple feature representations and multiple kernels. Therefore, the efficiency of MKL algorithms is not high enough.Since multiple kernel learning was proposed, a lot of researchers are focusing on finding out an efficient way to solve the MKL problems and improving the training efficiency of MKL algorithms. So far, the research on improving training efficiency of MKL algorithms mostly focuses on reducing the complexity of solving the objective function, other than reducing the training dataset. Therefore, it is meaningful to trying shrinking and optimizing the training dataset to simplify the MKL problems.In this paper, the principles and main thought of MKL is studied and the existing research achievement of improving the efficiency of MKL algorithms is given a detail introduction. After that, a novel MKL method based on cooperative clustering was proposed to improve the efficiency of MKL, which is called Multiple Kernel Learning based on Cooperative Clustering (CC-MKL). The method of cooperative clustering is a clustering algorithm based on κ-means algorithm and synergetics, which can find out the samples most likely to be chosen as the support vectors in on small scale. Using this method of cooperative clustering, the scale of training dataset will be reduced without losing classification information, which can significantly improve the efficiency of MKL algorithms with the condition of ensuring the classification accuracy change little. Experimental results show that Multiple Kernel Learning based on Cooperative Clustering can improve the training efficiency of MKL algorithms, especially for large-scale training datasets.In addition, when using multiple kernel support vector machine to deal with multiclass problems with the strategy of One-vs-All, it will cause the phenomenon of data imbalance. And it is also inefficient due to using all dataset for each binary classification. In this paper, cooperative clustering is also improved to preprocess the training dataset, which can make positive class and negative class become balance. The results of multiclass experiments show that using improved cooperative clustering for multiclass problems can significantly improve the training efficiency and the prediction accuracy.