Unified global adaptive anchor graph learning and factorization for incomplete multi-view clustering

Incomplete multi-view clustering (IMVC) aims to partition samples with partially observed heterogeneous views. Although anchor-graph and matrix-factorization IMVC methods improve scalability, most still adopt stage-wise pipelines that optimize anchor construction, graph learning, and representation learning separately. Such decoupling may cause a mismatch between structure modeling and clustering representation, especially under high missing ratios and heterogeneous noise. To address this issue, we propose a unified global adaptive anchor graph learning and factorization framework, termed UGA²GLF. The proposed method jointly learns view-specific anchors, a shared coherent anchor graph, view weights, and a graph-factorized clustering embedding in a single objective. Moreover, a missing-aware, view-wise self-rescaled adaptive loss is introduced for continuous sample-wise reweighting, suppressing noisy, outlying, and unreliable observations. An efficient MM-BCD solver is developed with a conservative acceptance rule for sparse simplex projection. Experiments on eight benchmarks show that UGA²GLF achieves consistently strong clustering accuracy and scalable runtime, especially under high missing ratios and on large-scale datasets.