通信学报 ›› 2021, Vol. 42 ›› Issue (10): 130-139.doi: 10.11959/j.issn.1000-436x.2021185

• 学术论文 • 上一篇    下一篇

融合注意力胶囊的深度因子分解机模型

顾亦然1,2, 姚朱鹏1, 杨海根3   

  1. 1 南京邮电大学自动化学院、人工智能学院,江苏 南京 210023
    2 南京邮电大学智慧校园研究中心,江苏 南京 210023
    3 南京邮电大学宽带无线通信技术教育部工程研究中心,江苏 南京 210003
  • 修回日期:2021-05-24 出版日期:2021-10-25 发布日期:2021-10-01
  • 作者简介:顾亦然(1972- ),女,江苏南京人,博士,南京邮电大学教授、硕士生导师,主要研究方向为复杂网络、大数据处理等
    姚朱鹏(1997- ),男,江苏苏州人,南京邮电大学硕士生,主要研究方向为推荐算法、自然语言处理等
    杨海根(1983- ),男,江苏南京人,博士,南京邮电大学副教授、硕士生导师,主要研究方向为无线通信、虚拟论证、虚拟设计等
  • 基金资助:
    国防基础科研基金资助项目(JCKY2019210B005);国防基础科研基金资助项目(JCKY2018204B025);国防基础科研基金资助项目(JCKY2017204B011);国防重大工程基金资助项目(ZQ2019D20401);装备发展部仿真预研课题(41401030301)

Deep factorization machine model based on attention capsule

Yiran GU1,2, Zhupeng YAO1, Haigen YANG3   

  1. 1 College of Automation &College of Artificial Intelligence, Nanjing University of Posts and Telecommunications, Nanjing 210023, China
    2 Center of Smart Campus Research, Nanjing University of Posts and Telecommunications, Nanjing 210023, China
    3 Center of Wider and Wireless Communication Technology, Ministry of Education, Nanjing University of Posts and Telecommunications, Nanjing 210003, China
  • Revised:2021-05-24 Online:2021-10-25 Published:2021-10-01
  • Supported by:
    The National Defense Basic Scientific Research Program of China(JCKY2019210B005);The National Defense Basic Scientific Research Program of China(JCKY2018204B025);The National Defense Basic Scientific Research Program of China(JCKY2017204B011);The Major National Defense Projects of China(ZQ2019D20401);The Simulation Pre Research Project of Equipment Development Department of China(41401030301)

摘要:

针对深度学习中推荐模型特征组合单一、消解大量有价值特征信息以及过拟合等问题,设计了一种新型的注意力得分机制——注意力胶囊,提出了一种融合注意力胶囊的深度因子分解机模型。基于DeepFM模型,将用户历史点击行为与候选物品进行权重计算,降低了无关特征对模型的影响,充分挖掘了不同历史行为对用户兴趣的差异性影响。训练过程中加入自适应正则化式,在不影响训练速度的前提下,有效地减少了过拟合。在2个公开数据集上进行对比实验,实验结果表明,所提模型相对于其他模型在损失函数和GAUC上均有明显提升。

关键词: 推荐模型, 深度学习, 注意力胶囊, 因子分解机, 自适应正则化

Abstract:

Aiming at the problems of single feature combination of recommendation model, resolution of a large amount of valuable feature information, and over-fitting in deep learning, a new attentional scoring mechanism called attention capsule was designed, and a deep factorization machine model based on attention capsule was proposed.Users’ historical clicking and candidate items were processed through weight calculation based on the DeepFM model, reducing the impact of irrelevant features on the model, and the differential impact of different historical behaviors on users’ interests was fully explored.The adaptive regularization formulation was added to the training, which effectively reduced over-fitting without affecting the training speed.The comparison test on two public data sets shows that the proposed model is significantly enhanced in loss function and GAUC compared to other models.

Key words: recommendation model, deep learning, attention capsule, factorization machine, adaptive regularization

中图分类号: 

No Suggested Reading articles found!