Multi-Level Confidence Learning for Trustworthy Multimodal Classification

With the rapid development of various data acquisition technologies, more and multimodal come into being. It is important to integrate different modalities which are with high-dimensional features for boosting final classification task. However, existing methods mainly focus on exploiting complementary information modalities, while ignoring learning confidence during fusion. In this paper, we propose a trustworthy network via multi-level learning, referred as MLCLNet. Considering that large number feature dimensions could not contribute performance but disturb discriminability samples, mechanism suppress some redundant features, well enhancing expression discriminative in each modality. order capture inherent sample structure implied modality, design graph convolutional branch learn corresponding preserved representation generate modal-specific initial labels. Since samples from should share consistent labels, cross-modal label fusion module deployed correlations modalities. addition, motivated ideally orthogonality fused matrix, loss supervise separable representations. To best our knowledge, MLCLNet first work integrates both label-level classification. Extensive experiments four medical datasets conducted validate superior when compared other state-of-the-art methods.