<p>Graph convolution network (GCN) is a promising deep learning method, mainly used to process graph data, such as social network, knowledge graph, etc. The basic idea of GCN is to use local information to update the node feature vectors, so as to obtain global information and realize graph-level tasks. When dealing with large-scale graphs, the computational and storage overhead of GCN are expensive, which may lead to a decline in generalization performance. Fortunately, formal concept analysis (FCA) as a data mining method based on lattice theory, is mainly used to extract the conceptual relationship between data. By classifying and clustering the data, a complete concept hierarchy can be established to help people better understand the internal relationship between data and provide strong support for subsequent data analysis. To this end, this paper proposes a formal concept-enhanced graph convolution network (FC-GCN), which uses the FCA methodology to preprocess large-scale graphs, extract graph data information and deploy the graph-related services. Technically, the maximal cliques and concept stability are used for feature update to implement downstream tasks such as node classification, link prediction, and community detection thereby overcoming the problems of graph convolution methods on large-scale graphs. The FC-GCN model is trained with <i>Cora</i> dataset, and the performance of model accuracy with different parameters is analyzed using node classification. Compared with baselines on real-world datasets, FC-GCN has better performance on vertice classification, accuracy improvements range from 4.2 to 6.1%.</p>