国家重点研发计划 (2022YFC3301004); 国家自然科学基金 (62172177); 中央高校基本科研业务费 (2022JYCXJJ034)
在当前数据来源多样化且人工标记难度大的现实生活中, 半监督场景下多视角数据的分类算法在各个领域中都具有重要的研究意义. 近年来, 基于图神经网络的半监督多视角分类算法研究已经取得了很大的进展. 但是现有的图神经网络算法大多是在分类阶段进行多视角互补信息的融合, 反而忽略了训练阶段同一样本不同视角间互补信息的交互. 针对上述问题, 提出半监督场景下多视角信息交互的图卷积神经网络算法MIGCN (multi-view interaction graph convolutional network). 该方法通过在不同视角上训练的图卷积层之间引入Transformer Encoder模块, 使得同一样本在训练阶段都可以通过注意力机制自适应的在不同视角间获取互补性信息, 进而加强自身的训练; 除此之外, 还通过引入一致性约束损失让不同视角最终特征表达的相似关系尽可能一样, 促使图卷积神经网络在分类阶段更加合理的利用多视角特征之间的一致性和互补性信息, 进一步提升多视角融合特征的鲁棒性. 最后, 在多个真实世界多视角数据集上的实验表明, 相比于基于图的半监督多视角分类模型, MIGCN可以更好地学习到多视角数据的本质特征, 进而提升半监督多视角分类的准确性.
In current real life where data sources are diverse, and manual labeling is difficult, semi-supervised multi-view classification algorithms have important research significance in various fields. In recent years, graph neural networks-based semi-supervised multi-view classification algorithms have achieved great progress. However, most of the existing graph neural networks carry out multi-view information fusion only in the classification stage, while neglecting the multi-view information interaction between the same sample during the training stage. To solve the above issue, this study proposes a model for semi-supervised classification, named multi-view interaction graph convolutional network (MIGCN). The Transformer Encoder module is introduced to the graph convolution layer trained on different views, which aims to adaptively acquire complementary information between different views for the same sample during the training stage. More importantly, the study introduces the consistency constraint loss to make the similar relationship of the final feature expressions of different views as similar as possible. This operation can make graph convolutional neural networks during the classification stage better utilize the consistency and complementarity information between different views reasonably, and then it can further improve the robust performance of the multi-view fusion feature. Extensive experiments on several real-world multi-view datasets demonstrate that compared with the graph-based semi-supervised multi-view classification model, MIGCN can better learn the essential features of multi-view data, thereby improving the accuracy of semi-supervised multi-view classification.