Bidirectional loss function for Label Enhancement and distribution learning

Label distribution learning (LDL) is an interpretable and general paradigm that has been applied in many real-world applications. In contrast to the simple logical vector single-label (SLL) multi-label (MLL), LDL assigns labels with a description degree each instance. practice, two challenges exist LDL, namely, how address dimensional gap problem during process of exactly recover label distributions from existing labels, i.e., Enhancement (LE). For most LE algorithms, fact dimension input matrix much higher than output one always ignored it typically leads reduction owing unidirectional projection. The valuable information hidden feature space lost mapping process. To this end, study considers bidirectional projections function which can be problems simultaneously. More specifically, novel loss not only errors generated projection into but also accounts for reconstruction back one. This aims potentially reconstruct data data. Therefore, expected obtain more accurate results. Experiments on several datasets are carried out demonstrate superiority proposed method both LDL. Specifically, BD-LE achieves optimal performance 85.71% cases renders sub-optimal 13.09% cases. BD-LDL ranks 1st 90.48% across six evaluation measurements. Compared baseline methods outperform best baselines over 7.38% 9.98% average respectively.