%A DONG Xin-yue, FAN Rui-dong, HOU Chen-ping %T Active Label Distribution Learning Based on Marginal Probability Distribution Matching %0 Journal Article %D 2020 %J Computer Science %R 10.11896/jsjkx.200700077 %P 190-197 %V 47 %N 9 %U {https://www.jsjkx.com/CN/abstract/article_19363.shtml} %8 2020-09-15 %X Label distribution learning (LDL)is a new learning paradigm for learning on instances labeled with label distribution,and has been successfully applied to real world scenes such as face age estimation,head pose estimation,and emotion recognition in recent years.In label distribution learning,enough data labeled by label distribution is needed when people train a model with good prediction performance.However,label distribution learning sometimes faces the dilemma that labeled data is insufficient and that marking enough label distribution data means high annotation cost.The Active label distribution learning based on marginal probability distribution matching (ADLD-MMD)algorithm is designed to solve the problem of high annotation cost for label distribution learning,by reducing the amount of labeled data required to train the model and reducing the annotation cost accor-dingly.The ALDL-MMD algorithm trains a linear regression model.While ensuring the minimum training error of the linear regression model,it learns a sparse vector that reflects that which instance in the unlabeled data set are selected,so that the data distribution of the training data set and unlabeled data set after instance selection is as similar as possible.We relax the vector for easy calculation.An effective method to optimize the objective function in ALDL-MMD is given,and proof for the convergence of ALDL-MMD is also provided.The experimental results on multiple label distribution data sets show that the ALDL-MMD algorithm is superior to the existing active example selection methods on the two evaluation measures of "Canberra Metric" (distance) and “Intersection” (similarity) to measure that what degree of the label distribution of the instance is accurate,which reflects its effectiveness in reducing annotation costs.