<p>Non-Fungible Tokens (NFTs) are unique digital assets authenticated via blockchain and have emerged as transformative instruments for representing ownership across diverse sectors such as art, gaming, and decentralized finance. The rapid expansion of the NFT market, however, brings critical challenges into establishing purchase recommendation systems that reconciliate artistic appeal with financial value, extract insight from sparse and noisy interaction data, and match user preferences to market performance. To tackle these challenges and assist users in identifying valuable NFTs to purchase, this study develops a recommendation system that employs Data Envelopment Analysis (DEA) to assess NFTs’ ability to convert intrinsic characteristics into market success. The DEA model generates market performance scores and establishes reference relationships among NFTs. Using DEA results with NFTs’ intrinsic attributes, our new recommendation system DEA-Enhanced Transformer Network (DETN) is then built upon three complementary graphs to capture multifaceted relationships: a DEA-reference-based graph that links each NFT to its performance benchmarks, a DEA-performance-cluster-based graph that connects NFTs in the same DEA performance cluster, and a trait-similarity graph that groups NFTs with similar traits. A temporal Transformer uses these graph embeddings, DEA scores, and each user’s interaction history to produce NFT purchase recommendations that simultaneously align with user preferences and exhibit strong market performance. Applications on CryptoPunks dataset demonstrate that our proposed DETN outperforms existing recommendation systems such as collaborative filtering, graph-based models, and Transformer-based models. By integrating DEA with deep learning methods, DETN enhances recommendation precision and supports better informed decision-making processes.</p>