基于评论文本情感注意力的推荐方法研究

Research on Recommendation Method Based on Sentimental Attention of Review Text

基于评论文本的深度学习推荐方法主要利用评论文本刻画用户和项目的特征信息,学习用户对项目的评分关系,提升推荐的性能。现有研究工作在提高推荐系统精度质量的同时,忽略了情感特征在评分预测中的可解释性贡献。针对此问题,考虑了评论文本以及情感倾向分别在用户和项目嵌入中的作用,提出了一种基于评论文本情感注意力的推荐方法(IncorRAS-Rec)。首先,通过卷积神经网络(CNN)处理用户和项目的评论集,对用户和项目的评论文本进行评论特征表示,并提取相关的用户特征信息和项目特征信息,进而结合用户对项目的评分偏好,学习用户和项目的评论情感特征表示;其次,基于注意力机制为用户和项目聚合了相关的评论情感特征信息,学习用户和项目的嵌入表示;最后,结合偏置信息,基于用户和项目的嵌入预测了用户对项目的评分。在亚马逊公开数据集上进行了实验比较和分析,对模型性能进行了有效性评估。实验结果表明:所提IncorRAS-Rec模型不仅在均方根误差(RMSE)和平均绝对误差(MAE)评价指标上的性能要优于其他传统方法,而且能够实现基于情感特征在评分预测方面的解释性作用。

Recommendation methods of deep learning-based on review text mainly means to describe the feature information of users and items by terms of review texts,by rating relationship between users and items to improve the recommendation performance.Existing studies ignore the interpretable contribution of sentimental features on the rating prediction.To solve this problem,by incorporating the roles of review text and sentimental polarity orientation in the embeddings of users and items,respectively,a sentimental attention recommendation method was proposed based on review text(IncorRAS-Rec);Firstly,CNN(convolutional neural network)was used to handle review sets for users and items,represent the review features of users and items,and obtain relevant users features and items features;Then,by combining users′rating preference for items,users and items embedding with reviews′sentimental features were learned.Secondly,by aggregating reviews′relevant sentimental feature information for users and items in terms of attention mechanism,the embeddings of users and items were learned;Finally,the users ratings on the items were predicted based on users and items embeddings together with their bias information.The experimental comparison and analysis were carried out on public Amazon datasets,to evaluates the effectiveness of the model performance.Experimental results showed that the proposed IncorRAS-Rec model not only could outperform other traditional methods in terms of RMSE(Root mean square error)and MAE(Mean absolute error)metrics,but also implement the explanatory role of sentimental features in rating prediction.