Graph Convolutional Networks (GCNs) have demonstrated considerable success in multi-view learning tasks. Most existing multi-view graph convolution methods rely on traditional \(\textit{k}\) -nearest neighbor ( \(\textit{k}\) NN) algorithms for constructing adjacency matrices. However, the application of a uniform k-value across different datasets may introduce noise, thereby affecting the performance of the model. In this paper, we propose a novel topology construction method based on granular-ball computing, which adaptively adjusts the topology structure in response to the characteristics of the dataset. This approach effectively mitigates the impact of \(\textit{k}\) -value noise and improves the representation of node information. Specifically, the granular-ball algorithm facilitates high-quality clustering, resulting in topology structures that encapsulate richer information. Such enriched topologies foster stronger connections, thereby enhancing the performance of subsequent graph convolution operations. Experimental results demonstrate that our proposed method outperforms existing multi-view semi-supervised classification techniques.