Scalable Label Distribution Learning (SLDL) is proposed for multi-label classification, where different labels are described as distributions in a latent space with asymmetric correlation, independent of the number of labels, resulting in little computational consumption.