基于图神经网络与多相似度融合的推荐系统性能研究

Enhancing Recommendation System Performance with Graph Neural Networks and Multi-Similarity Fusion

摘要: 随着数据规模和信息复杂性的增加，传统的推荐系统方法很难满足现代用户的个性化需求。图神经网络作为专门为处理图结构数据而构建的方法，其引入能使推荐系统更准确地分析复杂的用户行为并预测用户偏好，而多相似度方法能从不同角度度量用户或物品之间的相似性，有助于提升推荐精度。因此，本文通过引入动态融合权重α，有效整合了图神经网络与多相似度方法，以兼顾两种方法的优势，从而提升推荐性能。本实验在数据集MovieLens100k进行，通过RMSE、MAE、AUC和Precision四项指标来评估不同模型组合和动态融合权重α下的性能表现。实验结果表明，融合GNN与多相似度模型可显著提升推荐性能，其中以GraphSAGE融合Cosine + Jaccard相似度组合表现最优。此外，引入动态融合权重α实现了良好的调节效果，尤其在GraphSAGE模型中提升明显。

Abstract: With the increasing scale of data and complexity of information, traditional recommender systems struggle to meet modern users’ personalized demands. As a methodology specifically designed for graph-structured data, Graph Neural Networks (GNNs) enable recommender systems to analyze intricate user behaviors and predict preferences more accurately. Meanwhile, multi-similarity approaches measure user/item similarities from diverse perspectives, contributing to enhanced recommendation precision. To leverage the complementary strengths of both paradigms, this paper introduces a dynamic fusion weight α to effectively integrate GNNs and multi-similarity methods, thereby improving recommendation performance. Experiments conducted on the MovieLens-100k dataset evaluated performance under varying model combinations and α values using four metrics: RMSE, MAE, AUC, and Precision. Results demonstrate that fusing GNNs with multi-similarity models significantly enhances recommendation performance, with the GraphSAGE integrated with Cosine + Jaccard similarities achieving optimal results. Furthermore, the dynamic fusion weight α provides effective regulation, yielding particularly pronounced improvements in the GraphSAGE model.