Improvement of personalized recommendation service level of intelligent libraries based on artificial intelligence and big data mining technology

Abstract This study aims to improve the level of personalized recommendation services in intelligent libraries of colleges and universities and address the issue of insufficient recommendation accuracy in existing models caused by static user representation and data sparsity. It proposes a lightweight graph neural network recommendation model integrating the idea of Singular Value Decomposition++ (SVD++), namely SVD++-Lightweight Graph Convolutional Network (SVD-LGCN). Taking Lightweight Graph Convolutional Network (LightGCN) as its backbone, this model learns the global embedding representation containing high-order collaborative signals from the user-book interaction graph. Moreover, in the prediction layer, it innovatively draws on the idea of SVD + + and dynamically aggregates all historical interaction behaviors of users to generate an enhanced user preference representation, thereby achieving refined and dynamic characterization of user interests. Experimental results on the public Book-Crossing dataset show that the proposed model has superior performance. Compared with the LightGCN baseline model, SVD-LGCN achieves a relative performance improvement of 4.1% in NDCG@20, a core indicator for measuring ranking quality. Additionally, its Precision@5 is 0.413 and Recall@50 is 0.589. The research conclusion confirms that the proposed model provides an advanced technical solution with both theoretical innovation and practical value for constructing an efficient and accurate personalized recommendation system for intelligent libraries of colleges and universities.