基于评论文本图表示学习的推荐算法

基于卷积或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、非连续的依赖关系的捕捉能力有限。针对该问题，提出一种基于评论文本图表示学习的推荐算法RGP。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。针对图中的每个节点，使用基于连接关系的图注意力网络加权融合其邻点信息，利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从而得到整个图的表征。在此基础上，将基于用户和项目ID的嵌入表征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，以得到最终的推荐结果。实验结果表明，与NARRE、DAML等算法相比，RGP算法可有效提高推荐精度。

Recommendation Algorithm Based on Graph Representation Learning of Review Texts

Previous recommendation systems based on Convolutional Neural Network(CNN) or Recurrent Neural Network(RNN) mainly capture the local and consecutive dependency between neighboring words in review texts. Therefore,they may not be effective in capturing the long-term,global and non-consecutive dependency.To deal with the problem,a recommendation algorithm,RGP,is proposed based on graph representation learning of review texts.For each individual user/item,the method builds a specific review graph where nodes represent the review words,and edges describe the connections between words.For each node in the graph,a connection-based graph attention network is used to weight and fuse its neighboring information.Then an interaction-based attention mechanism is proposed to reweight the nodes in the graph,and the representation of all the nodes in the graph is fused to obtain the graph representation.On this basis,the embedded representations based on user and item ID are coupled with their review graph representations for input,and scored by a factorization machine to obtain the recommendation results.Experimental results show that compared with NARRE,DAML and other algorithms,the proposed algorithm effectively improves the recommendation accuracy.