改进型RBF神经网络的多标签算法研究

Multi-label Learning for Improved RBF Neural Networks

针对已有的RBF神经网络多标签算法未充分考虑多个样本标签之间的关联性,从而导致泛化性能受到一定影响的问题,研究分析了一种改进型RBF神经网络的多标签算法。该算法首先优化隐含层RBF神经网络基函数中心求取算法——k-均值聚类。采用AP聚类自动寻找k值以获得隐含层节点数目,并构造Huffman树来选取初始聚类中心以防k-均值聚类结果陷入局部最优。然后构造体现标签类之间信息的标签计数向量C,并将其与由优化k-均值聚类得到的聚类中心进行线性叠乘,进而改进RBF神经网络基函数中心,建立RBF神经网络。在公共多标签数据集emotion上的实验表明了该算法能够有效地进行多标签分类。

A modified multi-label radial basis function（RBF）neural network algorithm that can fully consider the relationship between numbers of sample labels was presented.This improved algorithm is based on the fact that ignoring the relevance between sample labels may cause potential performance loss.The modified algorithm first optimizes the RBF basis function center calculation algorithm in hidden layer,i.e.k-means clustering.AP clustering is used to automatically find kvalues to obtain the node number of hidden layer and a Huffman tree is constructed to select the initial cluster centers to prevent the k-means clustering results falling into local optimal.Then a label counting vector Cthat reflects the correlation between the labels is constructed,and it is linearly multiplied with the clustering centers which are obtained through k-means clustering optimization to optimize the RBF basis function center and establish RBF neural network.Experiments using the public multi-label emotion data sets demonstrate the effectiveness of the proposed algorithm.