| The application of multi-label data is must more extensive than the single-label data’s. Relatively, although the classification of multi-label data is more expensive and complex than single-label, its prediction accuracy is much lower than a single tag for the trained classifier. So it needs us to present classification algorithms with low cost and high accuracy rate for multi-label classification. In order to achieve such a goal, we use active learning in multi-label classification. To make our approach more widely used, the type of data will also be trained as map data. This paper introduces a method, learning local and global consistency (LLGC), which can train single-label classifier. Combined with the feature of multi-label data in graph, we propose learning local and global consistency for multi-label (ML-LLGC). We obtain a set of independent classifiers, each of which is independent, for getting the multi-label classifier we needed.In this paper, we use a tool, transductive Rademacher complexity, to associate with the generalization error of our classifier. Making our classification function to be variable, we can construct a model of transductive Rademacher complexity, which can be solve to build a relational expression of generalization error. Because of the relation between transductive Rademacher complexity and empirical transductive Rademacher complexity, we rebuild the relational expression between generalization error and empirical transductive Rademacher complexity. While we get the minimum empirical transductive Rademacher complexity, we get the minimum value of the generalization error bound. Therefore, we optimize the empirical transductive Rademacher complexity by using Jensen inequality, Cauchy-Schwarz inequality and sequential optimization algorithm. When the complexity is being minimized, we can obtain a set of unlabeled nodes which have the most information, and the nodes should be label by people, which can be used to train our classifier. The method, we propose above, is an active learning, which is an iterative procession. Ultimately, we will obtain a classifier which met our requirement of generalization error bound.In this paper, using a medical test data for our experimental data, we present a method for training multi-label classifier in graph and predict the labels of each data set in the test set, and we obtain low Hamming loss and one-error rate. This paper also experiment another method of multi-tag which can construct the most classic classifier, but unpractical. This method is mainly used for comparing with our method. The comparison shows that the multi-label classification we proposed is low cost and high accuracy. |
