Multi-view learning has received increasing attention in recent years due to its ability to leverage valuable patterns hidden in heterogeneous data sources. While existing studies have achieved encouraging results, especially those based on graph convolutional networks, they are still limited in their ability to fully exploit the connectivity relationships between samples and are susceptible to noise. To address the aforementioned limitations, we propose a framework called geometric localized graph convolutional network for multi-view semi-supervised classification. This framework utilizes a diffusion map to obtain the geometric structure of the feature space of multiple views and constructs a stable distance matrix that considers the local connectivity of nodes on the geometric structure. Additionally, we propose a truncated diffusion correlation function that maps the distance matrix of each view into correlations between samples to obtain a reliable sparse graph. To fuse the features, we use learnable weights to concatenate the coordinates of the geometric structure. Finally, we obtain a graph embedding of the fused feature and topology by using graph convolutional networks. Comprehensive experiments demonstrate the superiority of the proposed method over other state-of-the-art methods.