ScholarWorks@성균관대학교

초록

The user cold-start problem is a major challenge in item recommendation. Existing methods address this issue by utilizing additional user and item features or relying mainly on meta-learning (Few-shot learning). However, these approaches often require additional data or complex meta-training procedures. Another strategy is to construct a user intent from knowledge graphs, with methods typically falling into either path-based or propagation-based paradigms. Yet, in sparse data scenarios, path-based intent modeling fails to extract enough meaningful paths, while propagation-based methods amplify noise from limited neighbors and struggle to generalize beyond local contexts. In this paper, we propose a Conjunctive Query Embedding-based Recommender system, CQER, which models user intent as a logical query composed of relation-level paths. Rather than relying on paths aggregation or neighborhood propagation, CQER encodes intent via conjunctive query embeddings that capture uncertainty more effectively in sparse settings. This structure allows our model to make accurate predictions for users with only a few interactions, without requiring an additional adaptation step. Specifically, CQER supports K-shot recommendation (i.e., K=1 or K=5), leveraging a small number of user-item interactions to infer intent. Our method consistently outperforms existing approaches across multiple datasets and additionally provides high interpretability. Our code and datasets are publicly available at https://github.com/kjh9503/CQER.

키워드