| As an important generalization of nonnegative matrix factorization(NMF),hypergraph regularization nonnegative matrix factorization(HNMF)integrates the local geometric manifold information of the data,and can better describe and mine the nonlinear or multivariate relationships between data sample points,avoid the problem of low decomposition accuracy caused by the loss of original information.However,in practice,the original data has noise or outliers,and the pre-constructed hypergraph will follow the decomposition,which cannot accurately reflect the relationship between the sample points.In addition,in order to verify the feasibility and effectiveness of the new algorithm,clustering experiments are carried out on the image datasets COIL20 and Yale,and the clustering performance of the low-rank matrix factorization with adaptive hypergraph regularizer(LMFAHR)algorithm is compared with that of the five algorithms:K-means,principal component analysis(PCA),nonnegative matrix factorization(NMF),graph regularized nonnegative matrix factorization for data representation(GNMF),low-rank matrix factorization with adaptive graph regularizer(LMAGR).The final experimental results show that on the COIL20 dataset,the accuracy(ACC)and normalized mutual information(NMI)of evaluation indexand have improved0.66%～1.48%,0.19%～1.43%,respectively.On the Yale dataset,the accuracy and normalized mutual information have been improved0.01%～4.29%,0.3%～8.44%,respectively.Therefore,the new algorithm was feasible and effective. |
