摘要
基于评分矩阵的矩阵分解模型被广泛研究与应用,但是数据稀疏性问题严重制约了该模型的推荐效果.基于评论文本的推荐模型能够从文本信息中刻画用户偏好和商品特征,有效缓解了评分数据的稀疏性,但忽略了评分矩阵中用户和商品的潜在因子.为了进一步提高推荐质量,融合评分矩阵和评论文本的推荐模型被相继提出,但其仅仅局限在浅层线性特征层面,而且用户特征与商品的高级抽象特征未被充分挖掘,因此本文提出深度学习模型DeepCLFM(Deep Collaborative Latent Factor Model).该模型基于预训练的BERT模型,结合双向GRU和注意力机制从用户评论和商品评论中提取用户和商品的深层非线性特征向量,并根据用户和商品的编号映射出用户和商品的潜在隐向量.为了充分融合深层非线性特征和隐特征,DeepCLFM将用户和商品的深层特征向量与潜在隐向量以一、二阶特征项的方式产生深度特征项来预测出用户对商品的评分.在5组公开数据集上,以推荐结果的均方误差MSE作为评估指标进行对比实验,结果表明DeepCLFM的预测误差比多个优秀的基准算法更低,且平均预测误差最大降低了6.402%.
With the growing popularity of the Internet and smart mobile devices,people’s online time is rising.In order to improve office efficiency and consumption experience,the company provides a variety of products and services to meet the different needs of users,but it is also more difficult for users to quickly make satisfactory decisions from a large amount of information.Due to it can help different users to find out the items they are interested in through their historical behavior,the recommender system has become an extremely important part of online activities,such as online shopping,reading articles,and watching movies.To provide a personalized recommendation service,how to accurately predict the user’s rating of the item is a key issue that the recommender system needs to solve.Based on rating matrix,one of the most outstanding methods is matrix factorization,which has been widely studied and applied to model user preferences and item characteristics through rating data.However,the performance of these methods is severely restricted by the data sparsity problem,which can be seen as a phenomenon of the shortage of trainable data.To overcome this limitation,the recommendation models based on the review text can capture the user preferences and item features from the text data,effectively alleviating the sparsity of the rating data,but they ignore the latent factors of users and items in the rating matrix.With the comprehensive consideration of the above models and to further improve the recommendation quality,the model of combining rating matrix and review text has been proposed one after another.However,they are only limited to the linear latent feature level,in which the high-level abstract features of users and items fail to be fully explored.Therefore,this paper proposes deep learning model DeepCLFM(Deep Collaborative Latent Factor Model).First,the pre-trained BERT is used as the encoder of review text,which is a general-purpose "language understanding"model trained on a large text corpus like Wikipedia.Second,with the purpose of considering the latent relationship between different reviews in a review set,DeepCLFM extracts deep nonlinear feature vectors of users and items from review embeddings through a bidirectional GRU.Additionally,DeepCLFM introduces attention mechanism to measure the contribution of each review,and adopts matrix factorization module to learn latent factors according to the IDs of users and items.Finally,to fully integrate deep nonlinear features and latent factors,DeepCLFM generates deep interaction of them in the first and second order fashion to predict the user’s rating of the item.Experiments are conducted on five public datasets called Amazon Product Review,in which each sample contains user ID,item ID,user’s rating on the item(1~5 points),and user’s review text on the item.The mean square error(MSE)of the recommendation results is used as the evaluation metric.The results show that the prediction error of DeepCLFM is lower than that of many excellent benchmark algorithms,and the prediction error is reduced by a maximum of 6.402%.Moreover,DeepCLFM achieves a better performance than traditional matrix factorization in the"cold start"scenario.
作者
冯兴杰
曾云泽
FENG Xing-Jie;ZENG Yun-Ze(School of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300)
出处
《计算机学报》
EI
CSCD
北大核心
2020年第5期884-900,共17页
Chinese Journal of Computers
基金
国家自然科学基金委员会与中国民用航空局联合基金项目(U1233113,U1633110)
国家自然科学青年基金资助项目(61301245,61201414)资助。
关键词
推荐系统
评论文本
评分矩阵
神经网络
冷启动
recommender systems
review text
rating matrix
neural network
cold start
