부산대학교 도서관

Multimodal Collaborative Recommendation Systems (MCRecSys) have emerged as essential solutions in the digital age, addressing the challenges posed by a vast array of online content and services. This paper offers an overview of the landscape, emphasizing the significance of multimodal content while outlining primary approaches. We categorize these models based on their techniques, ranging from traditional matrix factorization to advanced deep learning configurations. A particular emphasis is placed on multimodal data fusion, highlighting three methodologies: early, intermediate, and late fusion. Additionally, we delve into evaluation metrics vital for assessing system performance and discuss commonly used loss functions such as mean square error (MSE), Bayesian Personalized Ranking (BPR) loss, and contrastive loss. Moreover, this study elucidates the role of self-supervised learning and pretraining in enhancing recommendation results. This study outlines a coherent pipeline, assisting researchers in grasping the prevalent concepts and methodologies in the field. At its core, the progression of MCRecSys underscores the potential of leveraging diverse modalities to deliver more precise and personalized recommendations in today's digital ecosystem.