E-MMKGR: A Unified Multimodal Knowledge Graph Framework for E-commerce Applications

Researchers Jiwoo Kang and Yeon-Chang Lee have introduced E-MMKGR, a novel framework designed to overcome the limitations of current multimodal recommender systems (MMRSs). These systems typically rely on a fixed set of item-side modalities (like images and text) and are built for specific tasks, which restricts their ability to incorporate new data types or generalize across different e-commerce functions. E-MMKGR addresses this by constructing a unified Multimodal Knowledge Graph (E-MMKG) specifically for e-commerce, creating a shared semantic foundation that can be applied to diverse applications like recommendation and search, moving beyond task-specific models.

The technical core of E-MMKGR involves using Graph Neural Network (GNN)-based propagation to learn unified item representations from the constructed knowledge graph, optimized through KG-oriented training. This approach allows the system to understand products through interconnected multimodal data rather than isolated features. In validation tests on real-world Amazon datasets, the framework demonstrated significant performance gains: up to a 10.18% improvement in Recall@10 for recommendation tasks and a substantial 21.72% advantage over standard vector-based retrieval methods for product search. These results highlight the framework's effectiveness and its key advantage of extensibility, meaning it can more easily incorporate new modalities (like 3D models or video) and adapt to new tasks without complete retraining, paving the way for more flexible and powerful e-commerce AI systems.