The invention provides a label distribution prediction method and device. The method comprises the steps that S1, acquiring a training set and initializing a potential factor matrix S2, constructing abasic objective function according to the training set and the potential factor matrix; S3, constructing a distance mapping function according to the potential factor matrix; S4, constructing an objective function according to the basic objective function and the distance mapping function; S5, enabling the objective function to solve a first-order partial derivative for the potential factor matrix; S6, determining a first factor matrix; S7, substituting the first factor matrix into the first-order partial guide quantity to obtain a step degree matrix; S8, judging whether the F norm of a gradient matrix is smaller than a critical threshold value or not, if so, executing S10, and if not, executing S9; S9, performing iterative optimization on the potential factor matrix to obtain a new firstfactor matrix, and executing S7; and S10, obtaining a label distribution vector of the to-be-predicted instance according to the first factor matrix. According to the scheme, the accuracy of label distribution prediction on the instance can be improved.