Abstract: Multiview clustering is an important research direction in data analysis, aiming to enhance clustering accuracy by integrating data from different perspectives. However, traditional multiview clustering methods improve clustering performance to some extent but often overlook the interaction and fusion of local and global features among views. Moreover, recent multiview deep clustering methods, which enhance representation capabilities through deep neural networks or contrastive learning, focus on local or global features, failing to integrate these two types of features within the same framework. To address these shortcomings, this paper proposes a deep multiview clustering model that integrates convolutional neural networks and Transformers (deep multi-view clustering network integrating local and global features, DMVCN-ILGF). The model designs parallel convolutional and Transformer branches to extract local and global features, respectively. To achieve effective feature fusion, the model introduces a feature interaction mechanism (FIM) and a feature fusion module (FFM), which fully integrates feature information from various views to enhance the interaction and fusion of different features, improving clustering performance. Additionally, instance-level and category-level contrastive losses are designed to compute the similarity between local and global features across views, optimizing the model??s representation capabilities and clustering outcomes. Experimental results demonstrate that the DMVCN-ILGF model significantly outperforms existing methods across multiple multiview datasets in clustering performance.

Key words: multi-view clustering, convolutional neural networks, Transformer, feature fusion

摘要： 多视图聚类是当前数据分析领域的一个重要研究方向，旨在通过整合来自不同视角的数据，提升聚类精度。然而，传统的多视图聚类方法虽然在一定程度上提高了聚类效果，但往往忽略了视图间局部与全局特征的交互与融合。此外，尽管近年提出的多视图深度聚类方法，通过深度神经网络或对比学习增强了表征能力，但大多只关注局部或全局特征，未能在同一框架下对这两类特征进行综合处理。针对这些不足，提出了一种融合卷积神经网络与Transformer的深度多视图聚类模型（DMVCN-ILGF）。该模型设计了并行的卷积分支和Transformer分支，分别用于提取局部特征和全局特征。为了实现特征的有效融合，引入了特征交互机制（FIM）和特征融合模块（FFM），通过充分整合各视图的特征信息，以增强不同特征的交互和融合，最终提升聚类性能。进一步地，还设计了实例级和类别级对比损失，分别计算各视图的局部与全局特征之间的相似性，从而优化模型的表征能力和聚类效果。实验结果表明，提出的DMVCN-ILGF模型在多个多视图数据集上均取得了显著优于现有方法的聚类性能。

关键词: 多视图聚类, 卷积神经网络, Transformer, 特征融合