A Comprehensive Survey on Multimodal Recommender Systems: Taxonomy, Evaluation, and Future Directions

Area of Science:

• Artificial Intelligence
• Computer Science
• Information Retrieval

Background:

• Recommender systems personalize user experiences using historical interaction data.
• Multimodal recommendation systems leverage diverse data types (e.g., text, images, audio) for improved accuracy.
• Existing systems often overlook complementary information present across different modalities.

Purpose of the Study:

• To provide a comprehensive review of recent research in multimodal recommendation.
• To delineate a standard pipeline, outline techniques, and classify existing models.
• To offer a performance analysis of baseline models and suggest future research directions.

Main Methods:

• Systematic literature review of multimodal recommendation research.
• Analysis of common pipeline stages and techniques used in multimodal systems.
• Performance evaluation of selected baseline models and development of a code framework.

Main Results:

• Identification of a common pipeline structure in multimodal recommendation systems.
• Classification of models based on their underlying methodologies.
• Performance insights into baseline models, highlighting factors influencing effectiveness.

Conclusions:

• Multimodal recommendation offers significant potential for enhanced personalization.
• The developed code framework facilitates understanding and implementation of state-of-the-art models.
• Open issues and future research directions in multimodal recommendation are identified.