基于信息选取的标记增强算法OA

To address the issue of label-irrelevant information impairing the accurate recovery of label distributions,this study proposes a label enhancement approach called label enhancement via informa-tion selection(LEIS).The approach works through the collaboration of an information selection network and a recovery network:the former integrates a feature reconstruction loss and a graph regular-ization loss to filter out irrelevant information and extract discriminative label-relevant features;the latter accurately reconstructs the label distribution based on these features by minimizing a distribution reconstruction loss.The entire model is trained end-to-end,with the label-relevant features acting as a coupling bridge,enabling joint optimization via gradient back-propagation to co-optimize the informa-tion selection and distribution recovery processes.Experimental results on multiple benchmark datasets confirm that LEIS demonstrates superior and more stable recovery performance compared to various state-of-the-art methods,verifying its effectiveness and competitiveness.