摘 要: 在当前的互联网营销环境中,多数模型尚未深入分析用户特征及用户行为的复杂性。对此,文章提出一种基于文本卷积神经网络(TextCNN)与多头注意力机制增强的xDeepFM(eXtreme Deep Factorization Machine)模型,即xDTCMAFM。首先,利用TextCNN高效地从文本数据中提取关键特征;其次,通过多头注意力机制进行不同子空间的特征提取;最后,使用xDeepFM模型实现深度显隐特征的交叉融合。实验表明,在两个互联网营销活动数据集上,该模型的AUC值分别达到了69.09%和72.98%,表现出了较好的性能,与xDeepFM等流行模型及融合注意力机制的改进模型相比均有一定提升。 |
关键词: 深度学习;多头注意力机制;TextCNN;xDeepFM;用户行为预测 |
中图分类号: TP183
文献标识码: A
|
基金项目: 浙江省重点研发“尖兵”攻关计划项目(2023C01119) |
|
Research on Participation Prediction of Internet Marketing Activities Based on TextCNN and Enhanced xDeepFM with Multi-head Attention Mechanism |
QIU Jiajie, HE Lili, ZHENG Junhong
|
(School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China)
qiujiajie1998918@163.com; llhe@zju.edu.cn; zdzhengjh@sohu.com
|
Abstract: In the current internet marketing environment, most models have not deeply analyzed the complexity of user characteristics and behaviors. In this regard, the paper proposes an xDTCMAFM model based on Text Convolutional Neural Network (TextCNN) and an enhanced eXtreme Deep Factorization Machine (xDeepFM) with a mult-i head attention mechanism. First, TextCNN efficiently extracts key features from textual data. Second, a multihead attention mechanism is employed for feature extraction in different subspaces. Finally, the xDeepFM model is used to achieve the cross-fusion of deep explicit and implicit features. Experiments show that on two internet marketing activity datasets, the proposed model achieves AUC values of 69.09% and 72.98% ,respectively, demonstrating better performance compared to popular models such as xDeepFM and the improved models that incorporate attention mechanisms. |
Keywords: deep learning; multi-head attention mechanism; TextCNN; xDeepFM; user behavior prediction |