Abstract
Deep neural networks have become a powerful technique for learningrepresentations from user-item interaction data in collaborative filtering (CF)for recommender systems. However, many existing methods heavily rely on uniqueuser and item IDs, which limits their ability to perform well in practicalzero-shot learning scenarios where sufficient training data may be unavailable.Inspired by the success of language models (LMs) and their stronggeneralization capabilities, a crucial question arises: How can we harness thepotential of language models to empower recommender systems and elevate itsgeneralization capabilities to new heights? In this study, we propose EasyRec -an effective and easy-to-use approach that seamlessly integrates text-basedsemantic understanding with collaborative signals. EasyRec employs atext-behavior alignment framework, which combines contrastive learning withcollaborative language model tuning, to ensure a strong alignment between thetext-enhanced semantic space and the collaborative behavior information.Extensive empirical evaluations across diverse real-world datasets demonstratethe superior performance of EasyRec compared to state-of-the-art alternativemodels, particularly in the challenging text-based zero-shot recommendationscenarios. Furthermore, the study highlights the potential of seamlesslyintegrating EasyRec as a plug-and-play component into text-enhancedcollaborative filtering frameworks, thereby empowering existing recommendersystems to elevate their recommendation performance and adapt to the evolvinguser preferences in dynamic environments. For better result reproducibility ofour EasyRec framework, the model implementation details, source code, anddatasets are available at the link: https://github.com/HKUDS/EasyRec.