Recommender systems predict users’ future interests primarily based on past interactions, limiting their capacity to discover novel interests. In this work, we investigate the task of User Novel Interest Discovery (UNID), aiming to predict users’ future interests in new item categories. Previous research has relied solely on user behaviors to predict novel items, neglecting the significant contributions of using additional knowledge, including common-sense knowledge, and structural and collaborative relationships among item categories. To address this, we propose MeRA, a Multi-faceted Relational Alignment framework for Large Language Models (LLMs) to integrate additional knowledge for UNID. MeRA leverages LLMs’ common-sense knowledge acquired in the pre-training stage and aligns LLMs with structural and collaborative relationships via two kinds of regularization. Extensive experiments validate the superiority of MeRA for UNID in terms of effectiveness and efficiency. We release code and data at https://github.com/GabyUSTC/MeRA .