Research Article Recommendation Using SVD++ in Academic Digital Libraries

Academic digital library recommender systems face a feedback asymmetry rarely as pronounced in other recommendation domains: explicit signals (bookmarks, saves, citation-derived ratings) are extremely sparse relative to implicit signals (views, abstract reads, downloads), since the overwhelming majority of a researcher's interaction with a digital library leaves no explicit trace of preference. SVD++ (Koren, 2008) extends matrix factorisation collaborative filtering by folding implicit feedback directly into the user latent-factor representation, augmenting the explicit-feedback factor vector with an aggregated implicit-feedback term rather than discarding the implicit signal as basic SVD does. This paper applies SVD++ to a multi-institutional academic digital library spanning 248,000 articles across five disciplines (Medicine: 74,400; Computer Science: 62,000; Social Sciences: 49,600; Physical Sciences/Engineering: 37,200; Humanities: 24,800) and 84,600 registered users (62,400 active, contributing a mean of 8.4 explicit interactions and 84 implicit interactions each), yielding a user-item interaction matrix with density of only 0.0034%. SVD++ achieves RMSE = 0.68, Precision@10=28.4%, Recall@10=38.4%, and NDCG@10=0.442 on a temporally held-out test set, outperforming basic SVD (RMSE 0.84, NDCG@10 0.324; +36.4% relative) by a margin that grows sharply with explicit-feedback sparsity: users in the sparsest quartile of explicit interactions see a 108.7% relative NDCG@10 improvement from incorporating implicit feedback, against only 11.7% for users in the densest quartile, directly confirming that implicit-feedback integration compensates for explicit-feedback scarcity rather than providing uniform benefit. For genuinely new users with zero explicit interactions, SVD++ achieves 18.4% Precision@10 against basic SVD's 7.2% (+155.6% relative), though new-item cold-start remains a distinct and only partially mitigated challenge. Deployed in production, SVD++ increases recommendation click-through rate from 7.6% to 12.4% (+63.2%) and doubles the rate at which recommended-and-downloaded articles are subsequently cited in users' own publications (4.2% to 8.4%), while diversity analysis reveals a measurable filter-bubble effect and a 3.4-fold over-representation of highly-cited articles in recommendations relative to the catalogue base rate.

Hali tarjima qilinmagan