摘要:
共享账户跨域序列推荐(shared-account cross-domain sequential recommendation, SCSR)是指在多个用户共同使用一个账户和用户的行为会在多个域中产生的情况下，给该账户推荐下一个可能会点击的项目.与传统的序列推荐任务相比，共享账户跨域序列推荐主要面临2方面的挑战：1)每一个账户里的交互行为是由多个用户产生的，并且这些用户的行为是混合在一起的；2)用户在1个域中产生的交互行为可能会提升推荐系统对该用户在其他域的推荐质量.目前，已有的一些相关工作大都是基于循环神经网络(recurrent neural network, RNN)的方法，但由于RNN本身固有的一些缺陷，导致基于RNN的方法不仅非常耗时，而且不能够很好地捕获交互行为之间的长期依赖关系.针对SCSR问题，提出了基于自注意力的跨域推荐模型(self-attention-based cross-domain recommendation model, SCRM)来解决这2个挑战.具体而言，首先引入1个多头自注意力网络来建模1个共享账户中多个用户参与的交互行为.然后，提出了一个基于多层交叉映射感知网络的跨域传输单元，以实现借助一个域的信息来提高另一个域的推荐质量.最后，通过一个混合推荐解码器整合了来自2个域的信息以实现在不同域中的推荐.在真实数据集HVIDEO上进行了实验，实验结果表明，与目前最新的基准方法相比，所提出的模型能在MRR和Recall这2个指标上取得了更加优异的结果；在运行效率上，比基于RNN的方法取得了更短的训练和学习时间.
Abstract:
Shared-account cross-domain sequential recommendation (SCSR) is the task of recommending next items in a particular context, where users share a single account, and their behavior records are available in multiple domains. Compared with traditional sequential recommendation tasks, SCSR is challenging due to: 1) The interactions generated by an account is a mixture of multiple users. 2) The behaviors in one domain might be helpful to improve recommendations in another domain. Recently，most of the related work is based on recurrent neural network(RNN). Due to the inherent drawbacks of RNN, RNN-based methods are time consuming and more importantly they fail to capture long-range dependencies of accounts’ interactions. In this work, we target at SCSR and propose a self-attention-based cross-domain recommendation model(SCRM) to address these two challenges. Specifically, to model the mixed interactions from multiple users of a single account, a multi-head self-attention network is first introduced. Then, to leverage the domain information in one domain to improve the recommendation in another domain, the cross-domain transfer network based on a multi-layer cross-map perceptual network is innovatively proposed. Finally, a hybrid recommendation decoder is explored to consider the information from both domains to achieve recommendation in each domain. We conduct experiments on a real-world dataset HVIDEO, and the experimental results show that SCRM outperforms all the baseline methods in terms of MRR and Recall. In terms of training efficiency, SCRM achieves shorter training and learning time than RNN-based methods.
百度学术
下载:
