Generative Incomplete Multiview Representation Learning With Learnable Graph.

Abstract

Incomplete and partially observed multiview data pose a fundamental challenge to representation learning, as missing views and highly complex cross-view inconsistencies hinder effective feature integration and alignment. While recent deep generative approaches have demonstrated strong potential for data imputation, their performance is frequently constrained by rigid graph assumptions or task-specific designs that limit adaptability. In this article, we propose a learnable graph-based generative representation learning framework that jointly models multiview dependencies and missing data through a learnable topological structure. The model captures both shared and view-specific relational patterns by adaptively fusing view-specific graphs into a unified structure. By integrating message propagation over adaptive graph structures with adversarial representation learning, the proposed model enables more reliable feature reconstruction and promotes consistent cross-view representations under incomplete settings. Extensive experiments on multiview semi-supervised classification benchmarks demonstrate that our method consistently outperforms existing approaches, validating its robustness and effectiveness in handling incomplete multiview scenarios.