Recommender systems (RSs) have become crucial parts in most web-scale applications, and the data sparsity is still one of the serious problems in RSs. Recently, the data sparsity can be tremendously alleviated by making use of informative data obtained from applications of edge devices and deep learning technologies with powerful data processing ability. In the face of extremely sparse rating data, the rich semantic information in reviews and the powerful feature extraction ability of convolutional neural network (CNN) contribute greatly to the improvement of recommendation performance. However, due to the complexity of natural language semantic components, it is common for many phrases in reviews to be separated by other words. Therefore, for the semantic information with unfixed intervals, the fixed geometric structure of CNN may lead to insufficient understanding of the user intention. Moreover, the usefulness of different reviews and the importance of different words in each review are various, which is vital for accurate modeling. In this paper, we propose a Double Attention-based Deformable Convolutional Network called DADCN for recommendation. In the proposed DADCN, two parallel deformable convolutional networks, which adopt the word-level and review-level attention mechanisms, are designed to flexibly extract features of both users and items from reviews. The parallel deformable convolutional networks jointly learn user preferences and item attributes, which is helpful to deepen the understanding of users’ attitudes. The word-level and review-level attention mechanisms are applied to intensify the critical words and informative reviews by assigning relatively high attention weights to them. Extensive experimental results on four real-world datasets demonstrate that the proposed DADCN outperforms four baseline methods.

推荐系统（RS）已成为大多数Web规模应用程序中的关键部分，而数据稀疏仍然是RS中的严重问题之一。最近，通过利用从边缘设备和具有强大数据处理能力的深度学习技术的应用程序中获得的信息性数据，可以极大地减轻数据稀疏性。面对评分数据极为稀疏的情况，评论中丰富的语义信息以及卷积神经网络（CNN）强大的特征提取能力极大地促进了推荐性能的提高。但是，由于自然语言语义成分的复杂性，评论中的许多短语通常被其他单词分隔开。因此，对于间隔不固定的语义信息，CNN的固定几何结构可能导致对用户意图的了解不足。此外，不同评论的有用性以及每个评论中不同单词的重要性各不相同，这对于准确建模至关重要。在本文中，我们提出了一个基于双重注意的可变形卷积网络DADCN进行推荐。在提出的DADCN中，设计了两个并行的可变形卷积网络，它们采用词级和评论级注意力机制，可以灵活地从评论中提取用户和项的特征。并行可变形卷积网络共同学习用户的喜好和项目属性，有助于加深对用户态度的理解。单词级别和评论级别的注意力机制通过分配较高的注意力权重来应用，以加强关键单词和信息性评论。在四个实际数据集上的大量实验结果表明，提出的DADCN优于四个基线方法。