摘要
针对深度多视子空间聚类网络在进行数据融合时不能区分各视图可靠性,以及缺乏对多视数据间一致性信息与互补性信息的利用等问题,提出一种基于自适应的权重融合深度多视子空间聚类(deep multi-view subspace clustering based on adaptive weight fusion,DMSC-AWF)方法。首先,通过使各视图共享同一个自表示层学习一个公共的表示矩阵,同时为各视图分别构建自表示层来学习各视图特定的表示矩阵,以此确保多视数据的一致性信息和互补性信息得以有效利用。然后,在共享自表示层基础上引入注意力模块来量化不同视图的重要性,注意力模块自适应地为每个视图数据分配权重。最后,在4个公开数据集上进行聚类实验,该方法的聚类结果相比于对比方法有明显的提升,并且,通过退化实验验证了注意力模块学习视权重的有效性和重要性。
In view of the inability of deep multi-view subspace clustering network to distinguish the reliability of each view when data fusion,and the lack of utilization of the consistent and complementary information between multi-view data,a deep multi-view subspace clustering method based on adaptive weight fusion(DMSC-AWF)was proposed.First,a common representation matrix was studied by making each view of sharing the same self-representation layer,and a self-representation layer was built for each visual to learn a specific representation matrix.The efficient use of consistent and complementary information that depends on the data was ensured.Second,the importance of different views was quantified by introducing attention modules based on the shared self-representation layer,which adaptively assigned weights to each visual data.Finally,clustering experiments were conducted on four public datasets,and the clustering results of this method were significantly improved compared with the comparison method.Moreover,the validity and importance of the attention module learning visual weight were verified by the degradation experiment.
作者
刘静
孙艳丰
胡永利
LIU Jing;SUN Yanfeng;HU Yongi(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)
出处
《北京工业大学学报》
CAS
CSCD
北大核心
2023年第7期758-768,共11页
Journal of Beijing University of Technology
基金
国家自然科学基金资助项目(61772048)。
关键词
深度子空间聚类
表示矩阵
多视
权重自适应
注意力模块
权重分配
deep subspace clustering
representation matrix
multi-view
weight adaption
attention module
weight distribution
