Label Distribution Learning by Exploiting Feature-Label Correlations Locally

Label distribution learning (LDL) is a novel learning paradigm that predicts the degree of representation of multiple labels to an instance. Existing algorithms use all features to predict label distribution. However, each label is often related to part of the features, hence considering other irrelevant features may lead to deviation in both instance searching and model prediction. In this paper, we propose a new LDL algorithm by exploiting the local correlation between features and labels (LDL-LCFL). The main idea is to exploit the local correlations between features and labels, which will be used in the improved k NN algorithm for prediction. Experiments were conducted on eight well-known label distribution data sets with four distance measurements and two similarity measurements. Results show that compared with nine popular LDL methods, our algorithm's prediction ranking is superior.