Modeling of Negative Feedback Refined via LLM in Recommender Systems

Most recommender systems have been proposed by utilizing the user feedback data, where implicit feedback is often used. However, in this setting, the dispreference signals in the original data get simply removed and treated the same as the unrated items. To address this limitation, recent studies have been proposed to utilize the negative feedback, and have shown how the recommender systems can be improved. One major limitation in these approaches is that the negative feedback data is generated through some fixed threshold. A recent work presented a new perspective by leveraging a Large Language Model (LLM) to flexibly distinguish positive and negative signals based on individual user intent. In this paper, we review this latest work, and extend it in three directions. First, we address the structural limitations in modeling negative feedback in the previous work. Second, we replace the existing loss function with a more robust approach, and show its effectiveness. Third, we propose a method for generating unified embeddings that integrate both positive and negative embeddings, thereby capturing user preferences and dispreferences in a holistic manner. Experimental results on five real-world datasets demonstrate that our model achieves state-of-the-art performance. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Chanwoo-Jeong-2000/ReFINe_plus</uri>