Research on the Application of Knowledge Graph-Driven Two-Dimensional Convolutional Embedding Methods in Recommender Systems

This article proposes a knowledge graph-driven recommendation method that integrates user preferences and contrastive learning to enhance the performance of the recommendation system. The study introduces knowledge-aware aggregation technology to extract user features from the user-project interaction graph, capturing preference characteristics using user-target factors. Additionally, a two-dimensional convolution method, incorporating both spatial and channel attention mechanisms, is employed to extract project features from the knowledge graph, with recommendation prediction achieved through the inner product. To address challenges posed by noisy knowledge graph data and the reliance on high-quality data, the paper introduces a model combining relationship learning with contrastive learning. This model leverages data augmentation techniques to generate subgraphs, enhancing the knowledge graph’s content. By comparing the subgraphs to the original graph, hierarchical features are learned to improve both the accuracy and robustness of the recommendations. The results show that the KG-UIR and KGCRL models significantly enhance recommendation performance and system robustness. They also help mitigate issues related to data sparsity and cold-start problems to some extent.