In many real-world applications, graph data often has missing attributes, is a challenging research task. Recently, attribute imputation methods based on multi-view networks have shown great potential in attribute-missing graphs. However, due to the missing attributes of certain nodes, existing methods for attribute-missing graphs can not effectively capture rich and complementary information between two views, thus limiting multi-view networks from learning high-quality attribute imputation. To address these problems, we propose a novel method named \(\textbf{M}\) ulti-view c \(\textbf{O}\) llaborative learning for graph attri \(\textbf{B}\) ute imput \(\textbf{A}\) tion(MOBA). Specifically, MOBA leverages a reliable augmentation strategy based on original graph relations, serving as a basis to aggregate attribute-observed neighboring node information. In the encoding stage, we introduce imbalanced encoders based on distinct propagation steps in different views, which effectively enhance the complementary information. Subsequently, to preserve more accurate node embeddings, MOBA introduces a multi-view collaborative learning strategy which aims to reduce the redundant information and maximize the consistency between two views. Extensive experiments on four benchmark datasets have demonstrated the effectiveness and superiority of our proposed MOBA over the state-of-the-art methods.
