基于评论特征提取和隐因子模型的评分预测推荐系统

评分预测是推荐系统研究的核心问题,通过用户的历史行为来预测用户对商品的评分,根据评分高低来推荐用户喜欢的商品.当前基于评论评分预测推荐系统普遍只使用卷积神经网络捕获局部特征或者循环神经网络捕获全局特征,忽略了将这两类特征的有效融合.针对现存问题,本文提出基于评论特征提取和隐因子模型的评分预测推荐模型,使用自适应感受野的卷积神经网络(CNN)提取局部特征,同时使用门控循环单元(GRU)提取全局特征,将不同特征融合为评论的嵌入表达.再结合隐因子模型(LFM)对用户的特征偏好和商品的特征属性进行建模.最后,通过对用户和商品的嵌入表达进行评分预测.实验结果表明,本文模型在5个数据集上均高于现有基线模型.

Rating prediction recommendation system based on reviews feature extraction and hidden factor model

Rating prediction is the core issue of the recommendation system research. It predicts the user''s rating of the product through the user''s historical behavior, and recommends the user''s favorite product based on the rating. The current recommendation system based on comment score prediction generally only uses convolutional neural network to capture local features or recurrent neural network to capture global features, ignoring the effective fusion of these two types of features. Aiming at the existing problems, this paper proposes a rating prediction recommendation model based on review feature extraction and hidden factor model, using adaptive receptive field convolutional neural network (CNN) to extract local features, and using gated recurrent unit (GRU) to extract global features. Fusion of different features into embedded representations of reviews. Then combined with the hidden factor model (LFM) to model the user''s feature preference and the feature attributes of the product. Finally, the rating prediction is made on the embedded representations of users and commodities. The experimental results show that the model in this paper is higher than the existing baseline model on the five data sets.