The development of Internet technology has produced a lot of high-dimension,varied and structure-complex data,including a large number of multi-label data that has single instances correspond to multiple labels.Most of the traditional supervised learning algorithms are used to process single-label data,and they cannot extract the internal information of multi-label data well.Therefore,multi-label learning has attracted the attention of scholars and has been studied by them.Multi-label data is also facing the problem of dimension disaster.To avoid the problem of over fitting and accuracy degradation caused by high dimension,this paper introduces NMF(Non-negative Matrix Factorization)and ELM(Extreme Learning Machine)to build a novel multi-label learning model.First,this paper studies the traditional NMF algorithm and propose FDAGNMF(Factorization Dimension Adaptive Graph Regularized Non-negative Matrix Factorization).Then optimize the number of hidden layer nodes,the center of hidden layers RBF function and the calculation method of output weights of ML-RBF(Mutil-label Learning Algorithm Based on Radial Basis Neural Network)algorithm,and propose FDAGNMF(Factorization Dimension Adaptive Graph Regularized Non-negative Matrix Factorization).Finally,a novel multi-label learning algorithm ensemble based on NMF and ELM are proposed by combining the two improved algorithms,and the effectiveness of the proposed algorithm is verified by experiments.The main contents and works of this paper are as follows:(1)Research on factorization dimension adaptive graph regularized non-negative matrix factorization algorithmThe traditional NMF need to determine the factorization dimension manually when it used to factor matrix.We use AP(affinity propagation)clustering algorithm to optimize the factorization dimension,and project the original data to the non-negative feature space.Then,we propose a factorization dimension adaptive non-negative matrix factorization algorithm(FDANMF)based on the AP clustering algorithm.To verify the effectiveness of the proposed algorithm,the KNN and ELM are used to classify the low dimension representations.To further improve the robustness of the low dimension representations of the original data,we propose a factorization dimension adaptive graph regularized non-negative matrix factorization algorithm(FDAGNMF)based on FDANMF and GNMF.The experimental results demonstrate that the classification accuracy of FDAGNMF algorithm combined with ELM and KNN classification algorithm is higher than that of FDANMF.(2)Research on a novel multi-label learning algorithm based on ELM and RBF neural networkFirst analyzes the defects of traditional ML-RBF algorithm in the process of multi-label learning.The number of hidden layer nodes are obtained by using k-means cluster algorithm to cluster the samples contained in each label.Among them,the K value is the ratio of the samples in each class,and they cannot reflact the true information of the samples.Therefore,the AP clustering algorithm is used to determine the number of the hidden layer nodes,and the output weight is calculated by using the RELM.The multi-label learning algorithm based on ELM and RBF neural network(ML-AP-RBF-RELM)is proposed.Finally,we compare the proposed algorithm with five multi-label learning algorithms on three multi-label data sets,and use five evaluation criteria to evaluate the effectiveness of the proposed algorithm.Further analysis the problems in ML-AP-RBF-RELM,we use AP clustering algorithm to determine the number of hidden layer nodes and the center of hidden layer RBF function in the same time,and use Lap-El M to calculate the output weights.We propose a multi-label learning algorithm ML-AP-RBF-Lap-ELM based on ELM and RBF neural network.And we compare ML-AP-RBF-Lap-ELM with ML-AP-RBF-RELM to verify the effectiveness of the proposed algorithm.(3)Reasearch on the multi-label learning algorithm based on NMF and ELMIn this paper,a multi-label learning algorithm,ML-NMF-Lap-ELM,which is based on the NMF and ELM,is proposed by combining the FDAGNMF and the ML-AP-RBF-Lap-ELM algorithm.First,the FDAGNMF algorithm is used to reduce the dimension of the the original data,and then the ML-AP-RBF-Lap-ELM algorithm is used to classify the low dimension representation the original data.The effectiveness of the ensemble algorithm is verified by comparing with the multi-label learning algorithm ML-AP-RBF-Lap-ELM on four multi-label data sets. |