计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (11): 1813-1827.DOI: 10.3778/j.issn.1673-9418.2006053

• 综述·探索 • 上一篇    下一篇

面向知识迁移的跨领域推荐算法研究进展

任豪,刘柏嵩,孙金杨   

  1. 宁波大学 信息科学与工程学院,浙江 宁波 315211
  • 出版日期:2020-11-01 发布日期:2020-11-09

Advances and Perspectives on Knowledge Transfer Based Cross-Domain Recom-mendation

REN Hao, LIU Baisong, SUN Jinyang   

  1. Faculty of Information Science and Engineering, Ningbo University, Ningbo, Zhejiang 315211, China
  • Online:2020-11-01 Published:2020-11-09

摘要:

由数据分布不均衡产生的数据稀疏和冷启动问题制约着个性化推荐系统进一步发展。随着迁移学习技术兴起,基于迁移学习的跨领域推荐为解决该类问题提供了可能。面向知识迁移的跨领域推荐算法通过迁移与目标领域不同但相关的辅助知识来解决目标领域中的推荐任务,提高目标域的推荐性能。而深度学习在非线性特征的学习和表示上的独特优势进一步提升了深度跨域推荐的算法性能。对近年来面向知识迁移的跨领域推荐算法展开综述,将当前主流算法分为传统跨领域推荐算法和深度跨领域推荐算法两大类,又按照应用的不同知识迁移技术对两大类跨域推荐算法分别梳理和阐述,从模型的可解释性、适用场景、用户特征的描述能力、模型评价等不同角度对当前各类跨域推荐算法做出深度分析和比较,总结其存在的问题和不足,探讨可能的解决方法,展望未来的研究趋势。

关键词: 数据稀疏, 迁移学习, 知识迁移, 跨领域推荐

Abstract:

Data sparseness and cold start problems caused by unbalanced data distribution restrict the further development of personalized recommendation systems. With the rise of transfer learning technology, cross-domain recommendation based on transfer learning provides possibility to solve such problems. This kind of algorithm can solve the recommendation task in the target domain by transferring appropriate auxiliary domain knowledge which is different but related to the target domain, and improve the performance of target recommendation task in the target domain. The unique advantage of deep learning in non-linear feature learning and representation has greatly improved the performance of deep cross-domain recommendation algorithms. A review of cross-domain recommendation algorithms for knowledge transfer in recent years is carried out. The state-of-the-art algorithms are divided into two categories: cross-domain recommendation and deep cross-domain recommendation. According to different knowledge transfer technologies, they are sorted and summarized separately. And then various algorithms are analyzed and compared in depth from different perspectives such as model interpretability, applicable scenarios, user characteristics, model evaluation, etc. Finally, this paper summarizes the existing problems and dificiencies of existing algorithms, explores passible solutions and forecasts the future development trend.

Key words: data sparsity, transfer learning, knowledge transfer, cross-domain recommendation