摘要
多视图聚类(MVC)作为无监督学习的一个关键研究方向,在处理大规模数据时,结合深度学习可以显示出其独特的优势。然而,三个关键挑战仍未得到解决:(1)如何有效地从海量数据中提取判别特征,同时减少过多的训练开销;(2)如何协调视图之间的异构特征表示;(3)如何保持多种模态之间的语义一致性。为了应对这些挑战,提出了一种基于线性注意力与簇对齐的框架(LRACA)。开发了一种标签驱动的锚点采样策略,其中类别感知K-means算法选择交叉视图对齐的锚点来指导训练方向的确定。为特征预训练设计了特定于视图的自编码器,并引入了一种动态低秩注意力机制,将键/值矩阵投影到线性子空间中,将注意力操作的计算复杂度从O(N^(2))降低到了O(N),同时显著增强特征区分性,并更新伪标签。提出了一种集群级对比学习范式,以聚类伪标签为纽带强化跨视图语义一致性,稳定和巩固了整个框架的聚类表现。对六个基准数据集的广泛实验表明,LRACA在聚类准确性、纯度和归一化互信息方面优于八个最先进的MVC基线,验证了其有效性和效率。
Multi-view clustering(MVC),as a pivotal research direction in unsupervised learning,when dealing with largescale data,incorporating deep learning can show its unique advantages.However,three critical challenges remain unresolved:(1)how to efficiently extract discriminative features from massive data while mitigating excessive training overhead,(2)how to harmonize heterogeneous feature representations across views,and(3)how to maintain semantic consistency among multiple modalities.To address these challenges,this paper proposes a deep multi-view clustering alignment guided by low-rank attention(LRACA).The technical contributions are threefold:firstly,a label-driven anchor sampling strategy is developed,where category-aware K-means algorithm selects cross-view aligned anchors to guide training direction determination.Secondly,this paper designs view-specific autoencoders for feature pretraining and introduces a low-rank attention mechanism that projects key/value matrices into linear subspaces,reducing the computational complexity of attention operations from O(N^(2))to O(N),while significantly enhancing feature discriminability and updating pseudo-labels.Thirdly,a cluster-level contrastive learning paradigm is proposed,leveraging clustering pseudo-labels as a bridge to strengthen cross-view semantic consistency,thereby stabilizing and consolidating the clustering performance of the entire framework.Extensive experiments on six benchmark datasets demonstrate that LRACA outperforms eight state-of-the-art MVC baselines in terms of clustering accuracy,purity and normalized mutual information,validating its effectiveness and efficiency.
作者
温珍平
孙颖慧
李杏峰
孙权森
WEN Zhenping;SUN Yinghui;LI Xingfeng;SUN Quansen(College of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;College of Computer Science and Engineering,Southeast University,Nanjing 210096,China)
出处
《计算机科学与探索》
北大核心
2026年第4期1103-1114,共12页
Journal of Frontiers of Computer Science and Technology
基金
国家自然科学基金(62372235,62406069)
中国博士后科学基金(2024M750425)。
关键词
多视图聚类
低秩注意力
对比学习
深度聚类
跨视图对齐
multi-view clustering
low-rank attention
contrastive learning
deep clustering
cross-view alignment
