作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (10): 405-417. doi: 10.19678/j.issn.1000-3428.0068484

• 开发研究与工程应用 • 上一篇    下一篇

基于用户多类型反馈行为序列的点击率预估模型

吴永庆*(), 王钰涵, 朱月   

  1. 辽宁工程技术大学软件学院, 辽宁 葫芦岛 125105
  • 收稿日期:2023-09-27 出版日期:2024-10-15 发布日期:2024-03-06
  • 通讯作者: 吴永庆
  • 基金资助:
    国家自然科学基金面上项目(52174184)

Click-Through Rate Estimation Model Based on User Multi-Type Feedback Behavior Sequences

WU Yongqing*(), WANG Yuhan, ZHU Yue   

  1. School of Software, Liaoning Technical University, Huludao 125105, Liaoning, China
  • Received:2023-09-27 Online:2024-10-15 Published:2024-03-06
  • Contact: WU Yongqing

摘要:

在推荐系统中, 现有的点击率预估模型通常采用用户近期点击过的行为序列作为模型的输入, 这将使模型难以得到全面的用户兴趣表示, 导致模型无法获得最优的精度。为了解决这个问题, 引入一个基于用户多类型反馈行为序列的点击率预估模型(UMFB)。该模型中多种类型的用户行为序列包括隐式反馈序列和显式反馈序列, 并且能够对不同的用户兴趣偏好进行建模。鉴于隐式反馈序列中包含大量的噪声, 结合基于傅里叶变换的兴趣去噪层来减轻干扰。此外, 为了解决显式反馈序列数据的稀疏性问题, 部署基于对比学习的兴趣增强层来提高建模效果。最后采用个性化兴趣融合层对用户的偏好进行建模。为了验证UMFB模型的有效性, 在短视频推荐领域的KuaiRand-Pure和KuaiRand-1K数据集上进行了对比实验, 结果表明, 与DMT基线模型相比, UMFB模型的AUC分别提高了1.07和0.91个百分点。

关键词: 推荐系统, 点击率预估, 行为序列建模, 多种行为序列, 对比学习

Abstract:

In recommendation systems, models for predicting click-through rates typically rely on a user's recent clicks as input. However, this approach can fall short in fully capturing user interests, limiting the model's accuracy. To address this issue, a new click-through rate estimation model based on User Multi-Type Feedback Behavior (UMFB) is developed, designed to handle various types of user feedback sequences. The UMFB model incorporates both implicit and explicit feedback sequences, enabling it to capture diverse user interest preferences. Given that implicit feedback sequences often contain significant noise, the study introduced an interest-denoising layer based on Fourier transform to reduce interference. Furthermore, to address the data sparsity issue in explicit feedback sequences, an interest enhancement layer based on contrastive learning is implemented to improve modeling accuracy. Finally, a personalized interest fusion layer is utilized to effectively model user preferences. To evaluate the effectiveness of the UMFB model, extensive experiments are conducted on the KuaiRand-Pure and KuaiRand-1K datasets in the context of short video recommendations. The results demonstrated that the UMFB model significantly outperformed other state-of-the-art baseline models, with an Area Under the receiver operator characteristic Curve (AUC) improvement of 1.07 and 0.91 percentage points on the respective real datasets.

Key words: recommendation system, click-through rate estimation, behavior sequence modeling, multi-type behavioral sequence, contrastive learning