评论感知的异构变分自编码器推荐模型

Review-aware heterogeneous variational autoencoder recommendation model

随着推荐系统的研究与发展,人们越来越关注个性化服务信息的准确推送,而对于推荐中数据稀疏的问题,传统评分信息协同推荐的方法很大程度上不能解决。因此人们将一些上下文信息引入到推荐系统中,而蕴含用户偏好的评论文本信息也被广泛用于缓解数据稀疏和冷启动的问题。自编码器作为一种无监督学习方法,在异常检测、人脸识别、数据增强和数据生成等领域具有优秀的表现,其中变分自编码器可以通过神经网络学习用户和项目潜在特征的分布。目前较少有研究利用用户评论信息融合的变分自编码器实现评论感知的推荐,该文提出一种评论感知的异构变分自编码推荐模型。首先,通过注意力机制和神经网络将评论上下文信息引入变分自编码器中,保留变分自编码器对评分信息潜在特征分布的学习,并在早期和后期两阶段进行特征融合,构建多模态的异构变分自编码器模型。其次,针对多模态模型训练,进一步优化引入复合先验项和平衡系数计算项。实验结果表明,该模型在召回率和归一化折损累计增益评价指标上都优于其他对比模型。

With the advancement of research and development in recommendation systems, more attention has been paid to the precise recommendation of personalized information. The traditional method of collaborative filtering cannot meet the demand due to the scarcity of data in recommendation;thus, contextual information has been introduced to recommendation systems. Review text information containing user preferences is also widely used to alleviate data sparseness and cold start problems. As an unsupervised learning method, the autoencoder performs well in anomaly detection, face recognition, data augmentation, and data generation. The variational autoencoder can learn the distribution of latent vectors of users and items via neural networks. At present, only a few researchers are working for review-aware recommendations using variational autoencoder. This paper proposes a review-aware heterogeneous variational autoencoder recommendation model that introduces comment context information into the variational autoencoder through attention mechanism and neural network. The learning about latent feature distribution of the rating information by the variational autoencoder is retained, and feature fusion is performed in the early and late stages to construct a multimodal heterogeneous variational autoencoder model. Besides, the compound prior term and the balance factor calculation term for multimodal model training are further optimized. The experimental results showed that the proposed model outperforms the state-of-the-art other baseline models in recall and the normalized cumulative gain.