Against the background of actively responding to population aging, the demand for community-embedded retirement building construction is increasing daily. Given the density and complexity of urban built-up areas, renovating existing community buildings in the context of inventory renewal has opened up a practicable development path for retirement building construction. However, the contradiction between the spatial structure of existing buildings and the functional requirements of retirement buildings has become a difficult problem. The environment of urban communities in China is diverse and complex, and the traditional program that relies solely on architects or community managers to carry out the renewal design has the shortcomings of low efficiency and low utilization rate. In this paper, we propose a spatial combination prediction method for retirement buildings based on a Graph Convolutional Neural network (GCN), which realizes the prediction of spatial combinations of retirement buildings by using the GCN node classification task model and the generative design method under the constraints of the existing site environment. This study also takes the building room characteristics and room connectivity into consideration, which can reflect the building usage more realistically in model training and generative design optimization, and this method has significant advantages in both time efficiency and space utilization.