计算机科学与探索 ›› 2024, Vol. 18 ›› Issue (7): 1792-1805.DOI: 10.3778/j.issn.1673-9418.2305022

• 理论·算法 • 上一篇    下一篇

融合多个性化桥和自监督学习的跨域推荐算法

王永贵,刘丹妮   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
  • 出版日期:2024-07-01 发布日期:2024-06-28

Cross-Domain Recommendation Algorithm Combining Multi-personalized Bridges and Self-supervised Learning

WANG Yonggui, LIU Danni   

  1. School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2024-07-01 Published:2024-06-28

摘要: 针对跨域推荐系统中目标域中项目交互较少的用户,提出一种融合多个性化桥和自监督学习的跨域推荐算法(MS-PTUPCDR)。首先,在目标域加入变分二部图编码器,使用变分推理框架生成潜在变量,目标域用户表示聚合其同构邻居信息。其次,将用户单一偏好桥扩展为用户多个性化偏好桥,将用户在多源域可转移的用户因子转移到目标域,在目标域加入多头注意力机制融合分别来自不同源域转换的用户潜在因子作为自监督学习的辅助任务。最后,在目标域中将聚合用户邻居因子和融合后的用户多源域转移用户因子进行自监督学习。在目标域通过用户自监督学习后的用户因子和目标域项目因子点积进行目标域项目评分预测。算法在Amazon和MovieLens两个数据集上进行实验,结果表明算法在MAE和RMSE两个评价指标上优于对比基线算法,在两个数据集上与最优对比基线算法相比,MAE平均提升1.96%,RMSE平均提升1.92%,验证了算法的有效性。

关键词: 跨域推荐, 用户多个性化偏好桥, 多头注意力机制, 自监督学习, 变分二部图编码器

Abstract: A cross-domain recommendation algorithm combining multi-personalized bridges and self-supervised learning (MS-PTUPCDR) is proposed for users with less project interaction in the target domain in the cross-domain recommendation system. Firstly, a variational bipartite graph encoder is added to the target domain, and a variational inference framework is used to generate potential variables. The target domain user representation aggregates their isomor-phic neighbor information. Secondly, the user??s single preference bridge is extended to the user??s multi-personalized preference bridge, the user's transferable user factors in the multi-source domain are transferred to the target domain, and the multi-head attention mechanism is added to the target domain to fuse the user's potential factors from different source domains as the auxiliary task of self-supervised learning. Finally, this paper aggregates user neighbor factors and the fused user multi-source domain transfer user factors for self-supervised learning. In the target domain, the project score of the target domain is predicted by the dot product of the user factor and the project factor of the target domain after the user's supervised learning. The algorithm is tested on two datasets, Amazon and MovieLens, and the results show that the algorithm outperforms the comparative baseline algorithm in terms of MAE and RMSE evaluation metrics. Compared with the optimal comparative baseline algorithm on both datasets, the MAE is improved by 1.96% on average, and the RMSE is improved by 1.92% on average, which verifies the effectiveness of the algorithm.

Key words: cross-domain recommendation, multi-personalized preference bridges for users, multi-head attention mechanism, self-supervised learning, variational bipartite graph encoder