A group recommendation method based on automatically integrating members' preferences via taking advantages of LLM

Abstract

Compared with personalized recommendation, group recommendation is more complex, achieving accurate recommendation that satisfy with all group members' preferences faces more severe challenges, including how to make a trade-off for the difference of preferences among group members, recommendation performance is easily affected by the problems of data sparsity and cold start, it is more difficult for users to understand the reasons for being recommended (i.e., poor interpretability), etc. Inspired by the strong text learning and understanding ability provided by large language models (LLMs), we propose a LLM-based group recommendation method for learning multi-view interaction topics of groups and items contained in various texts. This method can learn a group's preference by automatically integrating its members preferences without integrating policy, and analyze group/user preferences and understand group/user behaviors by using multi-view text mining. Specifically, in order to integrate rich group to item interaction information into the model, we designed a graph convolution network (GCN) model based on multi-topic learning, and denote the new model as topic-based graph convolution network via LLM (T-GCN-LLM). By applying graph convolutions on the multi-topic association graphs, the model can make a comprehensive representations for groups and users through using embeddings contained in multiple topics, so as to improve the group recommendations. We conducted extensive experiments on multiple real-world datasets to evaluate the T-GCN-LLM, the results demonstrate that our model can better represent the interactions between groups and items than many novel and high quality group recommendation methods. At the same time, the interpretability analysis experiment also proves the importance of incorporating the topics into the model to improve the interpretability of group recommendations.