电商推荐中知识图谱嵌入模型的比较研究OA

This study evaluated the performance of three typical knowledge graph embedding models,TransE,Rotate,and ComplEx,in e-commerce recommendation scenarios.Building a unified experimental platform based on the OpenBG500 dataset,conducting link prediction evaluation and segmenting relationship types for analysis.The experimental results showed that ComplEx performed the best overall(MRR 0.369,Hits@1 0.326)has outstanding advantages in modeling symmetric and asymmetric relationships;TransE training is efficient but has limited fine-grained modeling capabilities;Rotate is slightly better than ComplEx in a one to many relationship.This indicates that ComplEx is more suitable for multi relationship representation learning in e-commerce knowledge graphs,and can provide reference for cold start mitigation and long tail product support in recommendation systems.