Harnessing Multimodal Knowledge Graphs for Enhanced Recommender Systems: Insights from Graph Neural Networks and Attention Mechanisms

The explosive growth in the amount of information makes users consume information less efficiently, a phenomenon generally known as Information Overload. Recommender systems were created to solve such problems. Today, recommender systems are becoming the core technology of many Internet products. The core problem solved by recommendation algorithms is to replace the user to evaluate the goods or services that the user himself has not consumed and to recommend the goods or services that the user may be interested in to the user. To alleviate the problem of data sparsity, researchers have enhanced the characterization of items or users by introducing the Side Information of the items. In addition to the historical interaction information of the user and the item, the user and the item usually carry some auxiliary information, such as various attributes of the item and the basic information of the user. Reasonable utilization of auxiliary information can effectively alleviate such problems. This paper introduces a novel approach—a graph neural network based on knowledge graphs. This method uses a multimodal knowledge graph to assist recommender systems by converting the multimodal information fusion process in the knowledge graph into an information propagation process on the graph. The neighborhood information is better differentially aggregated through a trainable aggregation approach to obtain a more accurate representation of item node embeddings. In this paper, detailed experiments are conducted on two real recommender system datasets respectively, and the experimental results also show that our method outperforms the current mainstream recommendation algorithms.