Enhancing Anchor Representativeness and Stability in Scalable Multi-View Clustering with Prior Guidance

Anchor-based multi-view clustering has emerged as an effective paradigm for handling large-scale multi-view data, owing to its ability to reduce computational complexity by representing data samples via a small set of anchors. Existing methods typically fall into two categories: those using fixed anchors determined heuristically, and those that jointly learn anchors and anchor graphs. However, fixed-anchor methods lack adaptability and are sensitive to initialization, while learned-anchor methods risk overfitting to local data structures, resulting in poor generalization. In addition, most approaches rely on regularization terms to avoid trivial anchor graph solutions, which introduces additional hyperparameters and impairs stability. To overcome these limitations, we propose a Prior-Guided Anchor Learning method (PGAL) for scalable multi-view clustering. Specifically, we leverage pre-generated anchor priors to constrain the anchor updates, enhancing representativeness and mitigating overfitting. Meanwhile, the revised anchor graph module adopts a regularization-free quadratic formulation, enabling stable and unified consensus graph construction across views. We jointly optimize the anchors, anchor graph, view weights in a unified objective, and develop an efficient algorithm with theoretical guarantees. Experimental results on multiple benchmark datasets demonstrate that our method consistently achieves superior clustering performance compared to state-of-the-art alternatives. Our code is publicly available athttps://github.com/Tracesource/PGAL.