Interest in 3D modelling is growing, however, the retrieval results achieved for semantic-based 3D model retrieval systems have been disappointing. In this paper, we propose a novel semantic recommendation algorithm based on a deep belief network (DBN-SRA) to implement semantic retrieval with potential semantic correlations [between models] being achieved using deep learning from known model samples. The algorithm uses the feature correlation [between the models] as the conditions to enable semantic matching of 3D models to obtain the final recommended retrieval result. Our proposed approach has been shown to improve the effectiveness of 3D model retrieval in terms of both retrieval time and importantly accuracy. Additionally, our study and our reported results suggest that our posited approach will generalise to recommender systems in other domains which are characterised by multiple feature relationships.