Multi-view semi-supervised learning with adaptive graph fusion

Multi-view Semi-supervised Learning (MSL) is effective in using limited labels and considerable label-free data to improve learning performance. It has been successfully applied to a lot of real scenarios. In this study, we propose a model, termed MSL with Adaptive Graph Fusion (MSLAGF), which provides a novel solution for MSL. Unlike most existing methods propagating label information through the linear combination of pre-built fixed view-based similarity graphs, MSLAGF merges view-based graph construction, graph fusion, and label propagation. It adaptively learns view-specific graphs and automatically assigns weight coefficients to them. A multi-view fusion optimal graph is cleverly learned depending not only on the raw feature space but also on the dynamically predicted label space. Moreover, we present an efficient optimization algorithm to solve the formulated model. The view-specific graphs, the weight coefficients, the optimal graph, and the predicted labels are mutually negotiated and optimized in the optimization procedure. Extensive experimental results on six benchmark datasets validate the superiority.