Abstract:

In the collaborative filtering algorithm, the matrix factorization method based on rating data has been widely applied and developed, but the data sparsity problem affects the method recommendation quality. In view of this problem, a recommendation method (RHAOR) is proposed to integrate the review text hierarchical attention and outer product. Two parallel networks are used to process user review sets and item review sets, respectively. This paper applies aspect-level attention mechanism to the review text content, marks multiple words (or phrases) with aspect information, applies review-level attention mechanism to the review set, and marks valid reviews. The outer product is used to establish an outer product interaction matrix for user preferences and item features, and the multi-layer convolutional neural network is used to extract the outer product interaction feature. The outer product interaction feature is introduced into the improved latent factor model (LFM) for rating prediction. The experimental results show that the proposed method consistently outperforms traditional rating score and review based methods in root mean square error (RMSE) on Amazon and Yelp datasets.

Key words: collaborative filtering, data sparsity, review text, attention mechanism, outer product

摘要：

在协同过滤算法中，基于评分数据的矩阵分解方法得到广泛应用和发展，但评分数据稀疏性问题影响了该方法的推荐质量。针对此问题，提出一种联合评论文本层级注意力和外积的推荐方法（RHAOR）。采用两个并行网络，分别处理用户评论集和物品评论集。对评论文本的内容应用主题级注意力机制，标记多组带有主题信息的单词（或短语），对评论集应用评论级注意力机制，标记有效的评论。采用外积为用户偏好和物品特征建立外积交互矩阵，并对此矩阵采用多层卷积神经网络提取外积交互特征。将外积交互特征引入改进的潜在因子模型（LFM）中，进行评分预测。实验结果表明，在Amazon和Yelp数据集上，提出的方法在均方根误差（RMSE）上优于传统基于评分和评论的方法。

关键词: 协同过滤, 数据稀疏性, 评论文本, 注意力机制, 外积