多任务特征映射网络跨域推荐模型  

Multitask Features Mapping Network Model for Cross-domain Recommendation

作　　者：安秋雨 余艳 熊熙 AN Qiuyu;YU Yan;XIONG Xi(School of Cybersecurity,Chengdu University of Information Technology,Chengdu 610225,China;Advanced Cryptography and System Security Key Laboratory of Sichuan Province,Chengdu 610225,China)

机构地区：[1]成都信息工程大学网络空间安全学院,成都610225 [2]先进密码技术与系统安全四川省重点实验室,成都610225

出　　处：《哈尔滨理工大学学报》2024年第4期10-20,共11页Journal of Harbin University of Science and Technology

基　　金：国家自然科学基金(81901389);四川省科技计划项目(23ZDYF2088);教育部人文社会科学研究基金(22YJAZH120);成都信息工程大学青年科技创新计划(KYQN202227).

摘　　要：跨域推荐的关键在于如何有效的将用户特征从源域映射到目标域。以往的研究更关注用户特征的映射,忽略了映射前后用户特征的相似度以及物品特征映射的可能性;同时也只关注单一任务。因此针对以上问题提出了一种多任务特征映射网络跨域推荐模型(MTFMN)。该模型引入了用户特征映射网络,将用户在源域的特征映射到目标域,同时还引入了物品特征映射网络来辅助用户特征映射网络的学习;并在网络学习过程中使用欧氏距离来衡量映射前后用户和物品特征的相似度,以此作为网络学习过程中的一个优化策略;最后,利用映射后的用户特征和目标域上实际的物品特征完成偏好预测任务和评分预测任务。在Amazon数据集和豆瓣数据集上,MTFMN在评分预测任务的准确性上与主流模型相比有显著的提升。除此之外,模型还做了消融研究以证明模型中提出的物品特征映射网络和多任务优化策略的有效性。The key to cross-domain recommendation is how to effectively map user features from source domain to target domain.Previous studies focused more on mapping user features,ignored the similarity between user features and actual features after mapping and the possibility of item features mapping;it also only focused on a single task.Therefore,a multitask features mapping network model for cross-domain recommendation(MTFMN)is proposed to address the above issues.The model first introduces a user features mapping network,which can map the user features from source domain to target domain.It also introduces an item features mapping network to assist the learning of the user features mapping network.In the process of optimizing,Euclidean distance is used to measure the similarity between user and item features before and after mapping,as an optimization strategy in the process of training.Finally,preference prediction task and rating prediction task are performed on the user features obtained through user features mapping network and the actual item features in target domain.On the Amazon datasets and Douban datasets,MTFMN has significantly improved the accuracy of rating prediction tasks compared to mainstream models.In addition,the model also conducts ablation studies to prove the effectiveness of the item features mapping network and multitask optimization strategy proposed in the model.

关 键 词：推荐系统 跨域推荐 冷启动 数据挖掘 深度学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]