A Novel Label Enhancement Algorithm Based on Manifold Learning

We propose a label enhancement model to solve the multi-label learning (MLL) problem by using the in-cremental subspace learning to enrich the label space and to improve the ability of label recognition. In particular, we use the incremental estimation of the feature function representing the manifold structure to guide the construction of the label space and to transform the local topology from the feature space to the label space. First, we build a recursive form for incremental estimation of the feature function representing the feature space information. Second, the label propagation is used to obtain the hidden supervisory information of labels in the data. Finally, an enhanced maximum entropy model based on conditional random field is established as the objective, to obtain the predicted label distribution. The enriched label information in the manifold space obtained in first step and the estimated label distri-butions provided in second step are employed to train this enhanced maximum entropy model by a gradient-descent iterative optimization to obtain the label distribution predictor's parameters with en-hanced accuracy. We evaluate our method on 24 real-world datasets. Experimental results demonstrate that our label enhancement manifold learning model has advantages in predictive performance over the latest MLL methods. (c) 2022 Elsevier Ltd. All rights reserved.