Incomplete graph learning via data and representation-level interaction

Graph data is a fundamental resource for many modern applications. However, graph models may struggle when faced with incomplete data, such as incomplete structure or attributes. Despite partly solving the incomplete problem by modeling attributes and structure with a decoupled framework, existing methods ignore their interaction that may exploit the complementarity between attributes and structure, resulting in limited performance. In this work, we propose an interactive framework for incomplete graph learning on both the data and representation levels. Concretely, we design a data-level interactive completion approach that combines a decoupled preliminary completion with a fundamental interaction function realized by graph convolutional networks, effectively leveraging the complementarity between the completions of node attributes and graph structure to enhance incomplete graph completion. Then, we propose a representation-level interactive learning module with multi-view contrastive learning to handle representation-level interactions, generating better graph node representations and improving the performance of downstream tasks. Extensive experiments on eight benchmark datasets demonstrate the effectiveness of our methods in handling incomplete graphs, consistently outperforming existing state-of-the-art methods.