Abstract
Recommender systems play a pivotal role in personalizing user experiences by inferring preferences based on historical interactions with items. With the increasing prevalence of multimodal information, researchers have sought to develop recommender systems that can understand and interpret data across various modalities. These models can capture the hidden relations among various modalities and discover the complementary information that may be overlooked by recommender systems that exploit unimodal or user-item interaction data. This survey aims to provide a comprehensive review of the recent research efforts on multimodal recommendation. Specifically, we delineate a clear pipeline that is adopted by the majority of recommendation systems, outline commonly used techniques at each step of the pipeline, and classify existing models based on the methods used. We also conduct a performance analysis on several baseline models to identify the most well-performing models and to understand the factors that influence performance. Additionally, we design a code framework to assist new researchers in understanding the principles and techniques of this area. The code framework also serves as a platform for easily implementing the SOTA models and benchmarking the performance of recommender systems. Finally, we highlight some open issues and suggest potential research directions.