基于共享知识迁移学习的跨领域推荐研究

Cross-domain Recommendation Based on Shared Knowledge Transfer Learning

【目的/意义】基于迁移学习理论,提取多领域间共享知识模型,并进行有效的领域适应,提升目标领域推荐性能。【方法/过程】充分利用领域中的用户-物品评分矩阵,分别对多领域用户和物品进行潜在特征提取,并将用户-物品特征向量分别进行特征聚类;同时对多领域特征矩阵进行领域适应融合,得到共享知识模型;最后再运用迁移学习理论与方法,将收敛的共享知识模型从源领域迁移至目标领域,提高目标领域推荐性能。【结果/结论】实验结果表明,首先,多领域信息融合较于单领域推荐有着更好的推荐性能;其次,本文所提出的基于共享知识迁移学习的跨领域推荐模型在推荐效果上要优于当前业界已有的其它跨领域推荐算法。

【Purpose/significance】Based on the transfer learning theory, the shared knowledge model among multiple domains is extracted, and effective domain adaptation is carried out to improve the recommendation performance of the target domain.【Method/process】Fully utilize the user-item scoring matrix in the domain, perform potential feature extraction for multi-domain users and articles, and conduct feature-user clustering of feature-vectors. At the same time, the multi-domain feature matrix is adapted to the domain and the shared knowledge model is obtained. Finally, using the migration learning theory, the converged shared knowledge model is moved from the source domain to the target domain, and the recommended performance of the target domain is improved.【Result/conclusion】The experimental results show that, first of all, multi-domain information fusion has better performance than single-domain recommendation. Secondly, the cross-domain recommendation model based on shared knowledge transfer learning proposed in this paper is better than other existing ones in the industry.