基于层次标签结构的标记分布学习

Label Distribution Learning Based on Hierarchical Tag Structure

【目的】考虑到标记分布学习中标记之间具有层次结构关系,将层次标签结构引入标记分布学习,提升标记分布学习的效果。【方法】提出一种基于层次标签结构的标记分布学习算法(Hierarchy Label Distribution Learning Algorithm,H-LDL),根据样本在各层次的标签,利用条件概率描述粗、细两个层次的结构关系,并通过层次加权损失函数及其优化策略调节层次间标记的准确分布。【结果】在两个公开数据集上进行实验,用了5个指标进行效果检测,其中,BU_3DFE数据集在Euclidean、Squared、K-L指标中较基线算法最低值分别降低了3.99%、1.07%、3.10%,Intersec和Fidelity指标较基线算法最高值分别提升了4.24%、0.67%,COMP数据集在Euclidean指标上降低了0.48%,在Squared、K-L指标未见明显降低,在Intersec和Fidelity指标上提升了0.45%、0.02%。【局限】仅考虑了标签之间粗层次和细层次两层结构关系,当标签具有其他更复杂的层次结构关系时需进一步研究。【结论】加入层次标签结构后标记分布误差有明显减小,有效提升了标记分布学习的效果。

[Objective]This paper focuses on the complex hierarchical relationship between tokens in label distribution learning.It enhance performance by adding the hierarchical tag structure to the label distribution learning model.[Methods]We proposed a hierarchy-based label distribution learning algorithm(H-LDL),which used conditional probability to describe the extensive and intensive tag structural relationship.We also adjusted the exact distribution of each level by the function of hierarchical weighted loss and its optimization strategy.[Results]We examined the new model on two public datasets.The Euclidean,Squared,and K-L scores decreased by 3.99%,1.07%,and 3.10%on BU_3DFE dataset compared to the baseline model,while Intersec and Fidelity improved by 4.24%and 0.67%.On COMP dataset,the Euclidean decreased by 0.48%,but the Squared and K-L showed no significant decrease,while Intersect and Fidelity metrics increased by 0.45%and 0.02%.[Limitations]We only included two hierarchical relationships in the new model.Further research is needed for more complex hierarchical relationships.[Conclusions]A hierarchical label structure effectively improves the performance of label distribution learning.