期刊文献+

联合自监督学习强化的多行为多任务推荐算法

Multi-behavior multi-task recommendation algorithm integrating self-supervised learning enhancement
下载PDF
导出
摘要 为解决多行为推荐研究存在的未能全面捕获多行为交互特征,忽略点击等隐式反馈数据存在的大量噪声标签等问题,提出了联合自监督学习强化的多行为多任务推荐算法。首先,从行为影响权重和行为隐含语义两方面感知多行为交互特征,并将特征融合到嵌入传播过程,增强节点嵌入的表达能力;然后,构建自监督学习辅助任务,通过多视图对比学习避免模型对噪声标签过拟合;最后,联合有监督的多行为推荐任务和自监督学习辅助任务,采用多目标损失优化策略进行多任务学习,获取更加准确的用户、项目嵌入。通过实验分析表明,该算法在HR和NDCG指标上较对比算法均有一定提升,证明了算法的有效性和优越性。 To solve the problems of multi-behavior recommendation research such as failing to comprehensively capture multi-behavior interaction features and ignoring a large number of noise labels present in implicit feedback data such as clicks,this paper proposed a multi-behavior multi-task recommendation algorithm integrating self-supervised learning enhancement.Firstly,it sensed the multi-behavior interaction features from both behavior influence weights and behavior implicit semantics,and fused the features into the embedding propagation process to enhance the expressiveness of node embeddings.Then,it constructed the self-supervised learning assistance task to avoid model overfitting to noisy labels through multi-view comparison learning.Finally,it combined the supervised multi-behavior recommendation task the self-supervised learning assistance task and used a multi-objective loss optimization strategy for multi-task learning to obtain more accurate user and item embeddings.The experimental analysis shows that the algorithm has a certain improvement in both HR and NDCG indexes compared with the comparison algorithm,which proves the effectiveness and superiority of the algorithm.
作者 沈学利 张荣凯 Shen Xueli;Zhang Rongkai(College of Software,Liaoning Technical University,Huludao Liaoning 125105,China)
出处 《计算机应用研究》 CSCD 北大核心 2023年第9期2688-2693,共6页 Application Research of Computers
基金 国家自然科学基金资助项目(62173171)。
关键词 推荐系统 多行为推荐 自监督学习 多任务优化 recommender system multi-behavior recommendation self-supervised learning multi-task optimization
  • 相关文献

参考文献3

二级参考文献9

共引文献22

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部