融合复杂先验与注意力机制的变分自动编码器  

Variational Auto-Encoder Combining Complex Priori and Attention Mechanism

作　　者：沈学利 马玉营 梁振兴 SHEN Xueli;MA Yuying;LIANG Zhenxing(School of Software,Liaoning Technical University,Huludao,Liaoning 125105,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105

出　　处：《计算机工程》2022年第11期55-61,共7页Computer Engineering

基　　金：辽宁省教育厅科学技术项目(LJ2020FWL001)。

摘　　要：传统变分自动编码器模型通常使用标准正态分布作为隐向量先验,当应用于推荐系统等复杂任务时容易导致模型过度正则化和隐向量解耦表现不佳。融合复杂隐向量先验与注意力机制,建立变分自动编码器模型。使用多层神经网络生成的隐向量先验分布替代标准正态分布作为假设先验分布,使得模型能根据数据学习先验分布并获得更多的潜在表征。在单层隐向量的基础上添加辅助隐向量,联合辅助隐向量与数据特征向量再生成隐向量,增强了隐向量的低维表现能力和解耦性。借助注意力机制的特征信息选择特点,对隐向量中重要节点赋予更大的权重值,使其能传递更重要的信息。在数据集Movielens-1M、Movielens-Latest-Small、Movielens-20M和Netflix上的实验结果表明,该模型的Recall@20、Recall@50、NDCG@100相较于基线模型平均提升了12.95%、10.80%、10.48%,具有更高的推荐精确度。The traditional Variational Auto-Encoder(VAE)model typically uses the standard normal distribution as the implicit vector priori. When solving complex tasks such as the recommendation system,it easily leads to overregularization of the model and poor performance of implicit vector decoupling. A VAE model combining a complex implicit vector priori and attention mechanism is built to solve these problems.First,the implicit vector prior distribution generated by a multilayer neural network is used to replace the standard normal distribution as the hypothetical prior distribution so that the model can learn the most appropriate prior distribution based on the data and obtain more potential representations.Next,an auxiliary implicit vector is added to regenerate the auxiliary implicit and data feature vectors into an implicit vector based on the single-layer implicit vector. Compared with the original structure,it can significantly improve the low-dimensional representation ability and decoupling of the implicit vector.Finally,based on the feature information selection of the attention mechanism,the attention mechanism is added to two implicit vectors to increase the weights of the critical nodes so that the implicit vector can transmit more important information.Experiments were performed on the public datasets:Movielens-1M,Movielens-Latest-Small,Movielens-20M,and Netflix.The results show that the proposed model has better evaluation indexes than the experimental comparison model on the Recall@20,Recall@50,and [e-mail] average increase values are 12.95%,10.80%,and 10.48%,which show increased recommendation accuracy.

关 键 词：推荐系统 协同过滤 深度学习 变分自动编码器 辅助隐向量 复杂先验 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]