Method for enhancing label correlation based on global-local label specific feature learning

The invention discloses a method for enhancing label correlation based on global-local label specific feature learning, and the method depends on a sparsity hypothesis T1 of Lasso regression, and guarantees that a necessary label specific feature T2 is not lost. Using tag specific features as indicators T4 of strong correlation between related tags by applying constraints T6 to samples with high tag correlation; l2 regularization is introduced to learn global label correlation T3, and the difference between a prediction result and a real label is minimized by using a binary cross entropy and a Lasso loss function (T5, T1); iteration is carried out continuously until convergence, and an output result is a matrix W; and the test sample label set Y is obtained by multiplying the test sample X'by the regression coefficient matrix W. According to the method, missing label values are deduced and updated by adopting an iteration method, so that a complete label matrix is recovered, manual calculation and workload ar