Siamese network with squeeze-attention for incomplete multi-view multi-label classification

Abstract

Multi-view multi-label classification (MvMLC) has garnered significant interest because of its ability to handle complex datasets. However, the inherent complexity of real-world data often results in incomplete views and missing labels, which limit the richness of data and hinder the accurate association of features with their corresponding categories. Additionally, the MvMLC task is intricate due to the need for diverse views to coherently represent the same entity, thus demanding the creation of stable and consistent multi-view representations that can ensure a reliable feature alignment process across heterogeneous perspectives. To address these challenges, we propose a model based on a Siamese network with squeeze attention (SSA) for incomplete multi-view multi-label classification (iMvMLC). Specifically, to capture the shared semantic information across different views, we combine cross-view collaborative synthesis (CCS) and viewwise representation calibration (VRC) mechanisms. CCS enhances the semantic interaction between views by introducing directive blocks and stacked autoencoders on top of the Siamese network, thereby improving the ability to extract shared semantic representations. The VRC mechanism uses contrastive learning with positive and negative sample pairs to refine the shared semantic space, ensuring higher feature consistency and better alignment across views. Furthermore, considering the task-specific importance variation exhibited by each view, we apply the squeeze attention-weighted fusion (SWF) strategy, which performs feature dimensionality reduction to amplify the key characteristics from each view and enables the model to flexibly adjust the influence of each perspective. Extensive evaluations conducted across five datasets demonstrate that the SSA method outperforms many existing approaches.