Deep multi-view clustering based on global hybrid alignment with cross-contrastive learning

Multi-view clustering (MVC) plays a pivotal role in handling complex real-world data. However, conventional multi-view clustering methodologies are inclined to prioritize the uniformity across divergent data perspectives, frequently neglecting the distinctive information encapsulated within each individual view. Such methodologies risk the omission or forfeiture of view-specific information that could be advantageous for the clustering outcomes, which in turn impedes our profound comprehension of the data’s intricate architecture and underlying representations. To overcome this limitation, this paper introduces an innovative framework of global hybrid alignment and cross-contrastive learning (GHA-CL). Specifically, it adaptively learns and fuses features from different views through view-specific autoencoder networks and attention fusion layers to obtain a more comprehensive consensus representation. The uniqueness of the GHA-CL framework lies in its combination of a global hybrid alignment (GHA) strategy with cross-contrastive learning. This combination not only ensures the full utilization of view-specific information during the clustering process, but also further strengthens the common features across views, thereby achieving an effective balance between view consistency and specificity. Finally, we conducted thorough experiments on five real datasets to confirm the effectiveness of the GHA-CL framework. The experimental results demonstrate that the GHA-CL framework exhibits significant advantages in clustering performance on multi-view datasets. Our code is available on GHA-CL.