Integrating static and dynamic preferences via meta-learning for user cold-start recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The user cold-start problem is a central challenge in recommender systems, as existing methods often fail by modeling users’ static and dynamic preferences in isolation. We present the Meta-Learning for Dual-Stream Model (MetaDSM), a novel framework that addresses this by holistically integrating these preference signals. MetaDSM constructs a comprehensive user representation by capturing static preferences with a light graph convolution network (LightGCN) and dynamic patterns with a gated recurrent unit (GRU). These complementary streams are synergized through a hierarchical fusion architecture adaptively balanced by a gating mechanism. At its core, a model-agnostic meta-learning (MAML) framework enables rapid personalization from sparse user interactions. Experiments demonstrate MetaDSM’s superiority in the challenging user cold-start scenario, achieving an NDCG@10 of 0.4492 on the MovieLens-1M dataset and 0.4445 on the sparser Book-Crossing dataset. These results represent significant relative improvements of 4.7% and 69.5%, respectively, over the strongest baselines, confirming the efficacy of our dual-stream design and meta-learning framework. The source code for this project is publicly available at Github.