AMOSL: Adaptive Modality-wise Structure Learning in Multi-view Graph Neural Networks For Enhanced Unified Representation

While Multi-view Graph Neural Networks (MVGNNs) excel at leveraging diverse modalities for learning object representation, existing methods assume identical local topology structures across modalities that overlook real-world discrepancies. This leads MVGNNs straggles in modality fusion and representations denoising. To address these issues, we propose adaptive modality-wise structure learning (AMoSL). To enable efficient end-to-end training, we employ an efficient solution for the resulting complex bilevel optimization problem. The effectiveness of AMoSL is demonstrated by its ability to train more accurate graph classifiers on six benchmark datasets. NTRODUCTION A large amount of data networks exhibits a unique structure known as graph-structured data. Irregular, non-Euclidean data characterize this data type and are frequently found in areas such as recommendation systems [1], social media networks [2], knowledge graphs [3], and molecular structures [4]. The analysis of graph-structured data has garnered substantial attention for its inherent inductive and transductive properties, which enable relational reasoning among entities (nodes) and their connections (edges). Graph Neural Networks (GNNs), through a series of advancements [5][6][7][8][9][10][11][12][13], have shown promising improvement in studying graph-structured data to conduct various downstream graph-related tasks.