TARGET: Efficient and generalizable multi-view partial multi-label learning via anchor graphs and high-order tensor correlation

Multi-view partial multi-label learning (MVPML) addresses weakly supervised classification from multi-view data with ambiguous candidate annotations, in which the provided labels are only partially valid. However, prior studies largely depend on graph-based disambiguation and simple view aggregation, leaving high-order inter-view correlations and label dependencies insufficiently characterized, while sample-level graph construction introduces substantial computational overhead. To tackle these issues, we develop an efficient and generalizable MVPML framework based on anchor graphs and high-order tensor correlation modeling (TARGET), which simultaneously performs ambiguous-label disambiguation and accurate classifier learning. Specifically, we first impose tensor nuclear norm (TNN) regularization on the rotated principal classifier tensor, enabling higher-order cross-view correlations and label dependencies to be modeled in a shared low-rank tensor space. In addition, representative anchors are leveraged to replace sample-level graph construction and build anchor graphs from anchor-to-instance similarities, leading to efficient graph-based disambiguation with substantially reduced computational cost. Moreover, we develop an efficient alternating optimization algorithm to solve the resulting problem with closed-form updates or proximal operations. Finally, we establish generalization guarantees for the training and inference predictors, explicitly characterize how their complexity terms scale with the key problem dimensions, and show that, with the remaining quantities fixed, a lower tubal rank tightens the principal-classifier complexity term, thereby theoretically supporting the proposed low-rank modeling of cross-view consistency and label dependencies. Under matched assumptions and equal noisy-label-term coefficients, TARGET further admits a tighter overall Rademacher-complexity upper bound than the corresponding existing bound in a sufficiently low-tubal-rank regime. Experimental results on multiple datasets validate the effectiveness and efficiency of TARGET.