基于伪标签纠正的半监督深度子空间聚类

Semi-Supervised Deep Subspace Clustering Based on Pseudo-Label Correction

提出了一种基于伪标签纠正的半监督深度子空间聚类算法。首先利用少量已知样本标签,对分类层产生的不精确伪标签进行纠正,从而提高伪标签的精确性和稳定性;其次从已知样本标签中获得成对样本信息,通过对比学习对自表达系数矩阵进行约束来提高聚类的性能。在4个常用数据集上的实验证明,在最多50个已知样本标签的情况下,提出的子空间聚类算法性能优于目前先进的子空间聚类算法。

A semi-supervised deep subspace clustering algorithm based on pseudo-label correction is proposed.First,a small number of labeled samples are used to correct the imprecise pseudo-labels generated by the classification layer to increase the accuracy and stability of pseudo-labels.Then,the pairwise sample information is obtained from the labeled data,and the self-representation coefficient matrix is constrained by contrastive learning to improve the clustering performance.A test on 4 commonly used datasets demonstrates that the proposed subspace clustering algorithm outperforms the current state-of-the-art subspace clustering algorithms with no more than 50 labeled data points.