Abstract: Aiming at the situation that label distribution learning involves the feature correlation information of samples, and data may have abnormal and noise values, a model is established by combining the self-representation properties of samples and correlation between sample and label, and a label distribution learning algorithm (LRNSR-LDL) based on joint linear reconstruction and non-negative sparse representation is proposed. Firstly, the self-representation attribute of features is used to establish the linear relationship between the sample feature spaces, and obtain the feature similarity space after linear reconstruction; Then, using the correlation between feature and label, the label distribution is represented by the feature similarity space through non-negative sparse matrix decomposition, and optimization model is established with the loss function respectively; Finally, the l_2,1-norm constraint is introduced to reduce the adverse effects of outliers and increase the model's generalization ability. The proposed algorithm is compared with three existing label distribution learning algorithms on six real data sets, and five similarity and distance indices are respectively used for evaluation. The final experimental results show that the proposed LRNSR-LDL algorithm has its advantages.

Key words: label distribution learning, linear reconstruction, non-negative sparse representation, l_2,1-norm