《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (7): 2133-2139.DOI: 10.11772/j.issn.1001-9081.2022060802

• 人工智能 • 上一篇    

基于领域融合和时间权重的招工推荐模型

叶坤佩1,2, 熊熙1,2, 丁哲1,2()   

  1. 1.成都信息工程大学 网络空间安全学院, 成都 610225
    2.先进密码技术与系统安全四川省重点实验室(成都信息工程大学), 成都 610225
  • 收稿日期:2022-06-01 修回日期:2022-10-08 接受日期:2022-10-17 发布日期:2022-10-25 出版日期:2023-07-10
  • 通讯作者: 丁哲
  • 作者简介:叶坤佩(1998—),男,江苏泰兴人,硕士研究生,主要研究方向:自然语言处理、推荐系统;
    熊熙(1983—),男,四川成都人,副教授,博士,CCF会员,主要研究方向:信息抽取、自然语言处理、推荐系统;
    丁哲(1982—),男,四川成都人,讲师,博士,CCF会员,主要研究方向:数据挖掘。
  • 基金资助:
    国家自然科学基金资助项目(81901389);国家社会科学基金资助项目(19BGL123);四川省科技计划项目(2021JDRC0046)

Recruitment recommendation model based on field fusion and time weight

Kunpei YE1,2, Xi XIONG1,2, Zhe DING1,2()   

  1. 1.School of Cybersecurity,Chengdu University of Information Technology,Chengdu Sichuan 610225,China
    2.Advanced Cryptography and System Security Key Laboratory of Sichuan Province (Chengdu University of Information Technology),Chengdu Sichuan 610225,China
  • Received:2022-06-01 Revised:2022-10-08 Accepted:2022-10-17 Online:2022-10-25 Published:2023-07-10
  • Contact: Zhe DING
  • About author:YE Kunpei, born in 1998, M. S. candidate. His research interests include natural language processing, recommendation system.
    XIONG Xi, born in 1983, Ph. D., associate professor. His research interests include information extraction, natural language processing, recommendation system.
    DING Zhe, born in 1982, Ph. D., lecturer. His research interests include data mining.
  • Supported by:
    National Natural Science Foundation of China(81901389);National Social Science Foundation of China(19BGL123);Sichuan Science and Technology Program(2021JDRC0046)

摘要:

针对推荐系统使用嵌入层&多层感知机(Embedding&MLP)范式学习用户表示时强特征过拟合和弱特征欠拟合的问题,以及使用门控循环单元(GRU)学习用户兴趣时没有考虑到当前行为对用户最终兴趣的影响力会随时间推移逐渐减弱的问题,设计了一种基于领域融合和时间权重的招工推荐模型(RecRec)。首先,RecRec采用新的领域融合层来代替传统的串联层,而领域融合层在多域特征上表现出显著的优越性能。然后,RecRec在兴趣演化层将时间权重融入GRU,并提出时间戳门控循环单元(TSGRU),而TSGRU能更准确地学习用户兴趣。最终,RecRec通过预测用户拨通率来实现个性化推荐。实验结果表明,相较于YouTube DNN、Wide&Deep、融合注意力LSTM的协同过滤算法(ALAMF)和分期序列自注意力网络(LSSSAN),RecRec的AUC提高了0.03~0.36个百分点,说明RecRec能有效学习用户表示和用户兴趣。

关键词: 推荐系统, 深度学习, 注意力机制, 招工信息, 门控循环单元, 领域融合, 时间权重

Abstract:

To address the problem of strong feature overfitting and weak feature underfitting problem when learning user representations using Embedding layer & Multi-Layer Perceptron (Embedding&MLP) paradigm for recommendation systems and the problem of learning user interests using Gated Recurrent Unit (GRU) without considering that the influence of current behaviors on users’ final interests diminishes over time, a Recruitment Recommendation Model based on Field Fusion and Time Weight (RecRec) was proposed. In RecRec, firstly, a new domain fusion layer was adopted to replace the traditional tandem layer, and the domain fusion layer showed a significantly superior performance on multi-domain features. Then, time weight was incorporated into GRU in the interest evolution layer, and a TimeStamp Gated Recurrent Unit (TSGRU) was proposed, by which made the user interests were learned more accurately. Ultimately, personalised recommendations were achieved by predicting the dial-up rate of users. Experimental results show that the Area Under Curve (AUC) of RecRec improves by 0.03 to 0.36 percentage points compared to YouTube Deep Neural Network (DNN), Wide&Deep, Auxiliary LSTM-Attention Matrix Factorization (ALAMF) and Long-term & Short-term Sequential Self-Attention Network (LSSSAN), indicating that RecRec can effectively learn user representations and user interests.

Key words: recommendation system, deep learning, attention mechanism, recruitment information, Gated Recurrent Unit (GRU), field fusion, time weight

中图分类号: