基于HoFiBiAFM的点击率预测模型  

CLICK-THROUGH RATE PREDICTION MODEL BASED ON HOFIBIAFM

作　　者：马万民 王杉文 陈建林 牛浩青 欧鸥[1] Ma Wanmin;Wang Shanwen;Chen Jianlin;Niu Haoqing;Ou Ou(College of Computer and Network Security,Chengdu University of Technology,Chengdu 610059,Sichuan,China)

机构地区：[1]成都理工大学计算机与网络安全学院,四川成都610059

出　　处：《计算机应用与软件》2024年第10期170-176,241,共8页Computer Applications and Software

基　　金：国家重点研发计划项目(2018YFF01013304)。

摘　　要：在推荐系统中,FiBiNET、AFM等深度学习模型能够关注特征的重要性进行点击率预测。其中FiBiNET的深层模型使用DNN网络相当隐式地对特征交互进行建模,但是使用DNN学习高阶特征可能导致低阶特征交叉被稀释。通过叠加多层SENET注意力机制的方式学习高阶重要性特征,并加入高阶注意力分解机共同更新特征表示,构成一种新的点击率预测模型HoFiBiAFM。通过在Movielens-100K和Movielens-1M数据集上分别与其他CTR预测模型进行分类任务和回归任务的对比实验,结果验证了HoFiBiAFM模型的点击率预测效果。In the recommender system,deep learning models such as FiBiNET and AFM can focus on the importance of features for click-through rate prediction.FiBiNET s deep model uses DNN network to model feature interaction quite implicitly,but using DNN to learn higher-order features may lead to dilution of lower-order feature crossing.High-order important features were learned by superposing multiple SENET attention mechanisms and high-order attentive factorization machine was added to update feature representations.A new click rate prediction model HoFiBiAFM was formed.By comparing the classification task and regression task with other CTR prediction models on the Movielens-100K and Movielens-1M datasets,HoFiBiAFM s click-through prediction performance was verified.

关 键 词：推荐系统 点击率预测 特征重要性 SENET注意力机制 高阶注意力分解机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]