2026-10-14T20:18:49Z https://www.scidb.cn/oai

10.57760/sciencedb.hep.00002 2026-05-25T14:41:13Z

2026-05-25 Supplyment to the article"Relational contrastive multi-view clustering" doi:10.57760/sciencedb.hep.00002 en Supplyment to the article"In recent years, deep multi-view clustering (MVC) methods based on contrastive learning have received widespread attention. Although most existing contrastive MVCs have achieved remarkable progress (For details, see Online Resource 1), they still face two major challenges, as shown below. Inconsistent structural alignment: Recent advances in MVC have increasingly focused on achieving representation consistency across views [1–4]. For example, Chen et al. [5] explored the semantic information by contrasting the clustering assignments across multiple views to ensure consistent label representations. However, the above methods only focus on aligning instances themselves across either feature-level or label-level representations, while neglecting the consistency of structural relationships between different instances. For an image–text dataset, if two images are visually similar, their corresponding textual descriptions are also likely to exhibit semantic correlation. Nevertheless, many existing approaches fail to account for such cross-view structural consistency, which may cause semantic confusion under different views. Lack of global feature and semantic guidance: Most existing MVCs primarily depend on local representation alignment or pseudo-label supervision for optimization [6–9]. For instance, a selective alignment strategy is introduced [10], which aligns representations of different views at the sample level to boost clustering accuracy. While these approaches can enhance clustering performance, they typically lack a guidance mechanism that leverages global feature and semantic information, making it challenging to fully integrate the comprehensive information across multiple views. Contribution. To tackle the above problems, we propose a novel relational contrastive MVC (RCMVC) method, as shown in Fig. 1. The main contributions of this work are: 1) We propose a novel RCMVC method, achieving dual alignment at both the sample and relation levels. The former guides the representation learning for individual samples, while the latter captures relational consistency among different samples. 2) We introduce a relation-level contrastive module that explicitly aligns pairwise structural relationships across views. 3) Global feature information and semantic information are incorporated to guide local view learning." multi-view clustering; contrastive learning; structural alignment Mingliang XU Pei LV Xiaoheng JIANG Jin QIN Sijia LI Jiahao FAN Jinlan WANG Shizhe HU PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset http://www.doi.org/10.1007/s11704-026-51585-z Science Data Bank