Lifecycle-evolving recommendation with adaptive strategy transition for smart learning

Abstract

Educational data mining and learning analytics have become central to AI-enabled personalization in smart learning environments, where high-dimensional behavioral traces are continuously collected yet remain difficult to translate into reliable learning resource recommendations. Two long-standing barriers persist: cold-start prediction failures for newly enrolled learners with very limited behavioral history, and unreliable predictions caused by sparse user–learning resource interaction data . Existing studies typically address these issues in isolation and rarely embed the resulting algorithm within a deployable smart learning platform. We propose a lifecycle-aware recommendation paradigm that models strategy selection as a continuous function of user maturity, operationalized through an exponential decay transition mechanism. Within this paradigm, the cold-start regime employs naive Bayesian classification over directly observable behavioral features, while the mature regime deploys singular value decomposition with significance-weighted similarity. Experimental validation on the EdNet dataset shows that the framework achieves Precision@16 = 0.143, Recall@16 = 0.237, NDCG@16 = 0.202, and RMSE = 1.50, representing 4.4%, 16.7%, 4.7%, and 16.3% improvements over user-based collaborative filtering on Precision, Recall, NDCG, and RMSE respectively. The results demonstrate that combining educational data mining principles with lifecycle-aware strategy transition yields an interpretable and deployable solution for personalized learning resource recommendation in smart learning environments.