Prompt2Rec : Prompt based user and item Re-characterizing method for Recommendation

Seonjin Hwang, Younghoon Lee

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Review-based Recommendation System, Natural Language Processing, Prompt-based learning, Language Model, Review text

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: a novel user and item re-characterizing method for Recommendation, which introduces the Prompt-based learning paradigm of NLP

Abstract: Collaborative Filtering, which utilizes user-item interaction data is widely adopted in Recommendation Systems; however, the lack of interaction data can adversely affect recommendation performance. To address this issue, research incorporating Natural Language Processing (NLP) has made progress in leveraging review texts that contain rich information about user preferences and item attributes. Nevertheless, the conventional approach of integrating the entire review text and using it as an input, which has been widely used in previous research, can be vulnerable to noise (i.e., data with little relevance to user preferences or item attributes). In this study, we propose a novel user and item re-characterizing method called Prompt2Rec, which introduces the Prompt-based learning paradigm of NLP. It generates key factors that newly defined essential user and item characteristics from review texts and uses them as new information to train the recommendation model. Through experiments, we demonstrate that our proposed method can generate intuitive key factors related to user preferences and item attributes from reviews, and we validate that using these key factors in model training leads to improved performance compared to existing methods that rely on review texts. Furthermore, we explore the potential of visualizing the model's attention weights on the key factors for providing explanations of recommendations.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5282