Smart library personalized resource proactive recommendation system integrating user profiling and knowledge graphs.

The resource recommendation service of smart libraries is facing a dual dilemma of imprecise perception of user needs and passivity in recommendation patterns. Existing approaches mostly adopt classical paradigms of collaborative filtering or content-based filtering, exhibiting notable limitations in the following aspects: user profiles primarily rely on single-dimensional borrowing or retrieval behavioral data without integrating academic characteristics such as disciplinary backgrounds and research stages; knowledge graphs are mostly confined to collection metadata without embedding disciplinary ontologies and citation networks, limiting the mining of cross-domain resource semantic associations; and recommendation mechanisms lack context-aware proactive push capabilities, with recommendation timing being disconnected from readers' research rhythms and interest evolution trajectories. To address these issues, this paper proposes a smart library personalized resource proactive recommendation system that integrates multi-dimensional user profiling with a disciplinary knowledge graph. At the user modeling level, the system incorporates readers' borrowing records, retrieval logs, disciplinary backgrounds, and temporal behavioral patterns, employing a multi-head attention mechanism to achieve adaptive weighted fusion of multi-source heterogeneous features, and capturing the dynamic evolution trajectory of user interests through a gated recurrent unit. At the knowledge representation level, a library resource knowledge graph that fuses collection metadata, disciplinary ontologies, and citation networks is constructed, with a graph attention network employed for embedding learning of higher-order semantic relationships. Compared with GCN, GAT can adaptively assign differentiated aggregation weights to different neighbor nodes, making it more suitable for the library knowledge graph where different relation types contribute unevenly to recommendation. At the recommendation strategy level, a context-aware proactive recommendation mechanism based on interest drift detection is designed, which determines the optimal push timing by monitoring temporal changes in user profiles. The interest drift threshold is optimized through grid search on the validation set, and three types of contextual signals - access context, research stage, and interest drift - are fused through a gated aggregation mechanism to jointly trigger proactive push. Experimental results on a university library dataset demonstrate that the proposed method achieves 16.02% and 11.64% on Recall@10 and NDCG@10, respectively, representing improvements of 4.87% and 5.23% over the best-performing baseline methods. The push acceptance rate of proactive recommendations reaches 38.9%, significantly outperforming the random push strategy. Ablation experiments validate the effectiveness of each core module, with the knowledge graph module making the most prominent contribution to recommendation accuracy, primarily because it mines cross-domain resource correlations through disciplinary ontologies and citation networks that traditional metadata-based methods cannot capture.