Peishan Li Lutao Gao ^* Lilian Zhang ^* Lin Peng ^* Chunhui Bai Linnan Yang ^*

• Yunnan Agricultural University, Kunming, China

With the booming development of e-commerce, agricultural product recommendation plays an increasingly important role in helping consumers discover and select products. However, the following three problems still exist in the traditional agricultural product recommendation domain: (1) the problem of missing modalities made it difficult for consumers to intuitively and comprehensively understood the product information; (2) most of them relied on shallow information about the basic attributes of agricultural products and ignored the deeper associations among the products; (3) they ignored the deeper connections among individual users and the intrinsic associations between the user embedding and the localized user representation in different modalities, which affected the accuracy of user modeling and hindered the final recommendation effect. To address these problems, this paper innovatively proposed an agricultural product recommendation algorithm based on LLAVA and user behavioral characteristics, MP-LLaVRec(Modal Preference -Large Language and Vision Recommendation).It consisted of three main components: (1) LLAVA data enhancement, which introduced a multimodal macromodel to improve the understanding of node attributes; (2) agricultural product association relationship fusion, which constructed and improved the complex association network structure among products to ensure that the system can better understand the substitution relationship, complementary relationship, and implied consumption logic among products;(3) user modal preference feature extraction block, which deeply mined the interaction data between consumers and products, and advanced the effective user feature information from the correspondence between global user representations and local modal user representations.We conduct experiments on a real dataset from Amazon's large-scale e-commerce platform to verify the effectiveness of MP-LLAVRec. The experimental results of four metirs, NDCG@10, NDCG@20, Recall@10 and Recall@20, showed that the method has a better performance than the baseline model.