基于FGx_Deep算法的深度推荐  

Depth recommendation based on FGx_Deep algorithm

作　　者：余梦梦 孙自强 YU Meng-meng;SUN Zi-qiang(Key Laboratory of Advanced Control and Optimization for Chemical Processes of Ministry of Education,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学化工过程先进控制和优化技术教育部重点实验室,上海200237

出　　处：《计算机工程与设计》2020年第11期3204-3211,共8页Computer Engineering and Design

基　　金：中央高校基本科研业务费专项基金项目(222201917006)。

摘　　要：为在推荐系统中更好挖掘用户物品特征和理解用户需求,提出FGx_Deep算法。利用FGCNN网络以原始嵌入矩阵生成新特征,对原始特征和新特征进行拼接后输入到深度因子分解机(DeepFM)算法中,构建FG_DeepFM算法;基于FG_DeepFM算法与xDeepFM算法融合,提出FGx_Deep算法,解决FG_DeepFM算法只进行隐式特征元素级交互问题,挖掘显式特征向量级交互,实现端到端训练。将该算法应用到Movielens数据集上,实验结果表明,FGx_Deep算法相较现有的推荐算法,在评分预测推荐领域和Top-N推荐领域中,都有效提升了推荐准确性和泛化性。To better understand user requirements and project characteristics in the recommendation system,the FGx_Deep algorithm was proposed.To construct the FG_DeepFM algorithm,FGCNN network was used to generate new features from the original embedding matrix,and the original and new features were spliced and inputted into the DeepFM algorithm.Based on the fusion of FG_DeepFM algorithm and xDeepFM algorithm,the FGx_Deep algorithm was proposed to solve the problem of FG_DeepFM algorithm only performing implicit feature elements-level interaction,which was used to mine explicit feature level interaction and realize end-to-end training.Based on the Movielens dataset,experimental results show that the FGx_Deep methods can improve the recommendation effects and model generalization in both the scoring prediction recommendation field and the Top-N recommendation field.

关 键 词：嵌入矩阵 特征拼接 推荐算法 深度因子分解机 评分预测推荐 Top-N推荐 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]