【摘要】 针对子空间聚类在学习聚类的稀疏关系表示时不够精准、聚类误差大的问题，本文提出了使用随机位移和旋转等对原样本进行数据增强的算法。通过交替地使用增强样本来训练和优化网络，并更新样本的集群分配，从而学习稳健的稀疏关系表示；在微调阶段，损失函数中样本的目标都是将原样本分配到集群的中心，正确的分配有利于网络训练，而目标错误的样本会误导网络训练。新模型利用一种无需额外超参数的自适应权重学习，在每次迭代中优先选择易分类的样本，将集群边界附近的样本排除在训练之外，避免难分类样本产生误导性的记忆，从而提高泛化能力。算法在三个标准数据集上进行了实验，与五种典型的聚类算法相比较，证明了所提出算法的性能提升，消融研究和敏感性分析进一步说明了该算法的有效性。更多还原

【Abstract】 In order to solve the problem that subspace clustering is not accurate enough and the clustering error is large when learning sparse relation representation of cluster, this paper proposes an algorithm that uses random displacement and rotation. By iteratively training and optimizing the network using augmented samples and updating the cluster assignments, a robust sparse relationship representation is learned. During the fine-tuning stage, the objective of the loss function is to assign the original samples to the centers of the clusters. Correct assignment benefits network training, while misassigned samples can mislead the training process. The new model utilizes an adaptive weight learning approach without the need for additional hyperparameters. It prioritizes easily classifiable samples in each iteration, excluding samples near the cluster boundaries from training to avoid misleading memorization caused by difficult-to-classify samples, thus improving generalization ability. The algorithm is experimentally evaluated on three standard datasets and compared with five typical clustering algorithms, demonstrating performance improvement. Further ablation studies and sensitivity analysis provide additional evidence of the effectiveness of the proposed algorithm.更多还原