Improved RBF Neural Networks For Multi-Label Classification

Classification is a key problem in pattern recognition which can be divided into two categories: single-label classification and multi-label classification. On account of multi-label classification is quite widely used in the practical application, it gets the attention of the researchers.Multi-label learning refers to by training samples set of known labels, setting up a corresponding model, to classify label set of testing set. For existing multi-label algorithm based on RBF neural network did not adequately examine the association between a number of sample labels, resulting generalization performance affected in a certain degree, so an improved multi-label algorithm based on RBF neural network is proposed. Experiment on the public multi-label data sets confirmed the effective of the proposed algorithm.The work of this thesis includes the following parts:(1) The research status and research significance are briefly introduced, then the vision and disposed on multi-label study are summarized.(2) Studying the k-means clustering algorithm to improve hidden layer RBF basis function center. Using the AP automatic clustering to find k values as the hidden layer node number and Huffman tree, select the initial cluster centers to find the right values and avoid k means clustering result falling into local optimum algorithm for multi-label data experiments. Experiment result shows that the new method can effectively improve the ability of multi-label classification.(3) The new algorithm is designed: at first, a label counting vector C is generated from the correlation between more labels samples. And then with linear fold with RBF basis function center to form new RBF basis function, building RBF neural network for the solution from the input layer to output layer. The new algorithm has the following attractive properties: improves the recognition performance; expands data generalization performance.(4) Experiments using the public multi-label data sets demonstrate the effectiveness of the proposed algorithm.