计算机工程

• •    

基于评论文本图表示学习的推荐方法

  

  • 发布日期:2020-12-09

Recommendation Method Based on Graph Representation Learning of Review Text

  • Published:2020-12-09

摘要: 目前的基于卷积或者循环神经网络的推荐系统,主要捕捉评论文本中相邻词之间的局部和连续依赖关系,对长期、 全局和非连续的依赖关系的捕捉能力有限,无法同时考虑在不同评论中同一词的不同邻点集的全局和非连续依赖以及相邻词 之间的长期传播依赖关系。针对该问题,提出一种基于评论文本图表示学习的推荐方法。将每个用户或项目的评论文本表示 成图,图的节点为评论文本的词,图的边为词与词的连接关系。对图中的每个节点,使用基于连接关系的图注意力网络加权 融合其邻点信息。在此基础上,使用基于交互关系的注意力机制,对节点重新赋权,并加权融合图中所有节点的表征得到整 个图的表征。将基于用户和项目 ID 的嵌入表征以及其评论图表征耦合输入并采用因子分解机进行评分预测,得到推荐结果。 实验结果表明,和 NARRE 与 DAML 等算法相比,该算法可有效提高推荐的精度。

Abstract: Previous recommendation system based on Convolutional Neural Network (CNN) or Recurrent Neural Network (RNN) mainly capture the local and consecutive dependency between neighboring words for review text. Therefore, they may not be effective in capturing the long-term, global and non-consecutive dependency between words coherently. They can not consider the global and non-consecutive dependency of different neighbors of the same word in different reviews, as well as the long-term dependency based on propagation between adjacent words. Aiming at this problem, a recommendation method based on graph representation learning of review text is proposed. The method builds a specific review graph for each individual user/item, where nodes represent the review words and edges describe the connection types between words. For each node in graph, a connection based graph attention network is developed to aggregate its neighboring information. Moreover, an attention mechanism based on interaction is proposed to reweight the nodes in graph which are aggregated to obtain the graph representation. The representations based on user and item ID and their review graph representations are coupled and input and scored by a factorization machine to obtain the recommendation results. Experimental results show that the proposed algorithm effectively improves the accuracy of the recommendation compared with NARRE and DAML algorithms.