Incomplete multi-view semi-supervised classification via dual-graph structure and dual-contrastive completion

Abstract

In the real world, data often have multiple views, and learning from these multi-view data can improve the accuracy and robustness of classification models. The success of existing multi-view classification relies on a large amount of labeled and complete multi-view data. However, this is very difficult for practical applications due to data collection techniques failures and expensive labeling costs. To address this challenge, this paper proposes a new incomplete multi-view semi-supervised classification framework. Specifically, we first consider the potential graph structures among samples from specific views and a global view separately, aiming to extract specific information from each view and integrate complementary information across views. Then, we designed a completion module based on view-specific graphs and dual-contrastive learning. This module completes missing views of samples based on the spatial similarity relationships among samples in view-specific graphs, and then enhances the discriminative ability of the completion using dual contrastive learning. Finally, a classifier based on a global-specific graph and graph convolutional network (GCN) is designed to classify unlabeled samples by using spatial relationships among all samples and scarce labels. Extensive experiments, including comparison with existing algorithms, visualization analysis, ablation experiments, and parameter sensitivity examination, are conducted on seven real world datasets to showcase the effectiveness of the proposed framework.