| Multi-view data are data captured from different patterns,sources,spaces,and other forms with similar high-level semantics.In recent years,an increasing number of researchers have used multi-view learning to capture shared information from multiple views and thus improve classification performance.However,most existing multi-view classification models the lack of information sharing among different views from the representation,fusion,and decision levels are not uniformly addressed,which leads to limited performance improvement of the models.Therefore,how to integrate the representation,fusion and decision layers into a unified framework becomes a problem to be solved.In this paper,multi-view classification is investigated based on multi-view gated information allocation and mutual correlation analysis,and the main contributions work is as follows:(1)We propose a new nonlinear transformation method MVGID(Multi-View Gated Information Distributor)for multi-view data,which can better capture the important information of different views,change the output behavior of the views,and reduce the problem of missing features during the nonlinear transformation.(2)We propose a new method for multi-view information fusion by incorporating MVGID and cross-correlation operations operations into the network.It enables sufficient information sharing among views and focuses on the potential importance of single views.(3)Based on the above two methods,we propose a new Gating Cross-Correlation Network(GCCNet)model for multi-view classification,which is based on multi-view gated information distribution and cross-correlation analysis.GCCNet extracts cross-view complementary information by automatically extracting features and discarding noisy view information and reduces the negative impact of low-quality views on decision making by applying an adaptive weighted joint decision strategy.We conducted extensive experiments on 6 real datasets with hand-crafted pre-processed features and 3 real datasets with CNN pre-processed features to evaluate the proposed model in comparison with state-of-the-art multi-view classification models.The results show that the model performs significantly better than the existing models on 8 benchmark datasets.The experimental results demonstrate the rationality and effectiveness of the proposed method.(4)We further migrated GCCNet to the molecular property prediction task and proposed Drug-GCCNet.Drug-GCCNet enabled the integration of multiple molecular representations and extraction of complementary information to improve the accuracy of molecular property prediction.We evaluated the performance of Drug-GCCNet on2 benchmark datasets and verified the effectiveness and generalization ability of GCCNet in this domain. |
