基于DeepFM的深度兴趣因子分解机网络  

Deep Interest Factorization Machine Network Based on DeepFM

作　　者：王瑞平 贾真[1] 刘畅[1] 陈泽威 李天瑞[1] WANG Rui-ping;JIA Zhen;LIU Chang;CHEN Ze-wei;LI Tian-rui(School of Information Science and Technology,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学信息科学与技术学院,成都611756

出　　处：《计算机科学》2021年第1期226-232,共7页Computer Science

基　　金：国家重点研发计划(2017YFB1401400)。

摘　　要：推荐系统能够根据用户的喜好从海量信息中筛选出其可能感兴趣的信息并进行排序展示。随着深度学习在多个研究领域取得了良好的效果,其也开始应用于推荐系统。目前基于深度学习的推荐排序算法常采用Embedding&MLP模式,只能获得高阶的特征交互。为了解决该问题,DeepFM在上述模式中加入了因子分解机(Factorization Machine,FM),能够实现端到端的低阶与高阶特征交互学习,但其缺乏用户兴趣多样性的表示。鉴于此,通过将多头注意力机制引入DeepFM,提出了深度兴趣因子分解机网络(Deep Interest Factorization Machine Network,DIFMN)。DIFMN能够根据待推荐的不同物品自适应地学习用户表示,展示用户兴趣的多样性。此外,该模型根据用户历史行为的种类添加了喜好表征,从而不仅能够应用于只记录用户爱好的历史行为的任务,还可以处理同时记录用户喜欢与不喜欢的历史行为的任务。采用tensorflow-gpu进行算法的实现,在Amazon(Electronics)和movieLen-20m两个公开数据集上进行对比测试,实验表明所提算法相比DeepFM分别有17.70%和35.24%的RelaImpr提升,验证了其可行性与有效性。The recommendation system can sort out and display the information that may be of interest from the mass of information according to users’preferences.As deep learning has achieved good results in multiple research fields,it has also begun to be applied to recommendation systems.However,the current recommendation ranking algorithms based on deep learning often use Embedding&MLP mode and can only obtain high-level feature interactions.In order to solve the problem that only high-order feature interaction can be obtained,DeepFM adds FM to the above mode,which can learn the low-order and high-order feature interaction end-to-end.But the DeepFM cannot express the diversity of user interests.In view of this,this paper proposes a Deep Interest Factorization Machine Network(DIFMN)by introducing the multi-head attention mechanism into DeepFM.DIFMN can adaptively learn the user representation according to the different items to be recommended,showing the diversity of user intere-sts.In addition,the model adds preference representations according to the type of user's historical behaviors,so that it can be applied not only to tasks that record only historical behaviors that the user likes,but also to tasks that record both historical beha-viors that the user likes and dislikes.This paper uses tensorflow-gpu to implement the algorithm,and performs comparative tests on two public datasets of Amazon(Electronics)and movieLen-20 m.Experiment results show that RelaImpr improves by 17.70%and 35.24%respectively compared to DeepFM,which validates the feasibility and effectiveness of the proposed method.

关 键 词：推荐算法 DeepFM 多头注意力机制 深度学习 CTR预测 用户兴趣建模 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]