Multi-view multi-dimensional collaborative completion for node classification with missing attributes

Abstract Multi-view learning improves node classification with missing attributes by mining complementary information from diverse views. Existing multi-view self-supervised completion approaches, which rely on graph autoencoders to reconstruct missing node attributes from available features and graph structures in an unsupervised manner, suffer from two main limitations: they often overlook inter-layer correlations within the encoder and excessively depend on the accuracy of the input graph structure, thereby restricting node classification performance. To address these issues, this paper proposes a multi-view multi-dimensional collaborative completion method (MV-MDCC) for node classification with missing attributes. Our approach first constructs two more disturbance-resistant views by integrating feature propagation with a personalized PageRank engine. Subsequently, a triple-constraint strategy is introduced to maximize cross-view consistency across three dimensions—inter-layer, attribute, and structure—enabling more precise representation alignment. Finally, a dual-decoder strategy is designed, comprising an attribute decoder and an edge decoder, to reconstruct node features and topological structure, respectively. The reconstruction constraints facilitate learning representations with heterogeneous complementarity, thereby boosting node classification accuracy. Extensive experiments demonstrate that MV-MDCC consistently outperforms existing methods on four benchmark datasets under various attribute missing rates, confirming its effectiveness and robustness.