基于特征交叉注意力网络的序列推荐算法

Feature cross aware attention network for sequential recommendation

为解决现有序列推荐算法只从项目级别序列中挖掘用户兴趣,并未探究项目属性及其交互对用户兴趣影响的问题,提出一种基于特征交叉注意力网络的序列推荐模型。通过构建项目属性级别注意力和序列级别注意力,更好挖掘用户兴趣;项目属性级别注意力旨在学习项目及项目属性间的自适应相关性;序列级别注意力聚焦从项目级别序列和属性级别序列上学习序列动态性。在两个公开数据集上的实验结果表明,所提方法相比其它主流序列推荐算法在Hit、NDCG和MRR指标上有明显提升。

To solve the problem that existing sequence recommendation algorithms only mine user interests from item-level sequences and do not explore the impact of item attributes and their interactions on user interests,a feature cross aware attention network was proposed.Item attributes were leveraged into user’s interest by developing the horizontal-level attention and vertical-level attention.The horizontal level attention focused on learning adaptive correlation among item and its attributes.The vertical-level attention learned sequential dynamic both on item-level sequence and attribute-level sequence.Experimental results show that Hit,NDCG and MRR of this method are significantly improved compared to that of other mainstream recommendation algorithms on two different datasets.