您的位置:首页  > 论文页面

基于注意力的深度因子分解机模型研究

发表时间:2019-06-28  浏览量:1643  下载量:265
全部作者: 温瑶瑶,吴为民,张丽艳
作者单位: 北京交通大学计算机与信息技术学院;重庆能源职业学院汽车与信息工程系
摘 要: 点击率(click-through rate,CTR)预估是计算广告学中的核心研究问题,现有研究方法构建的特征信息有限,并且无法区分不同特征信息对结果的影响。基于上述问题,本文提出基于注意力的深度因子分解机(attention-based neural factorization machine,ANFM)。首先,ANFM将低阶交叉特征输入到深度神经网络来自动学习丰富的高阶交叉特征,旨在减轻深度神经网络的学习压力,同时采用注意力机制捕获不同特征交叉对结果的贡献度。通过在两个实际数据集上的实验表明,ANFM模型在CTR性能上相较于因子分解机(factorization machine,FM)分别有5.9%和10%的提升,相较于Wide & Deep分别有1.9%和6.9%的提升。同时相较于其他深度点击预测模型,ANFM结构简单,性能更好。
关 键 词: 人工智能;点击率预测;特征交叉;低阶特征;高阶特征;注意力机制
Title: Research on attention-based neural factorization machine model
Author: WEN Yaoyao, WU Weiming, ZHANG Liyan
Organization: School of Computer and Information Technology, Beijing Jiaotong University; School of Automotive and Information Engineering, Chongqing Energy College
Abstract: Click-through rate (CTR) prediction is the core research problem in computing advertising. The existing research methods have limited feature information and cannot distinguish the influence of different feature information on the results. Based on the above problems, an attention-based neural factorization machine (ANFM) is proposed in this paper. Firstly, ANFM inputs low-order interactive features into deep neural networks to automatically learn rich high-order interactive features, aiming to reduce the learning pressure of deep neural network, and to use attention mechanism to capture the contribution of different feature interactions to the results. Experiments on the two actual data sets show that the ANFM model has 5.9% and 10% improvement respectively in CTR performance compared to factorization machine (FM), and has 1.9% and 6.9% improvement respectively compared to Wide & Deep. At the same time, compared to other deep click prediction models, ANFM has simple structure and better performance.
Key words: artificial intelligence; click-through rate; feature interaction; low-order features; high-order features; attention mechanism
发表期数: 2019年6月第3期
引用格式: 温瑶瑶,吴为民,张丽艳. 基于注意力的深度因子分解机模型研究[J]. 中国科技论文在线精品论文,2019,12(3):371-382.
 
8 评论数 0
暂无评论
友情链接