基于注意力交叉的点击率预测算法  

PREDICTION ALGORITHM OF CLICK THROUGH RATE BASED ON ATTENTION MECHANISM

作　　者：杜博亚 杨卫东[1] Du Boya;Yang Weidong(School of Computer Science,Fudan University,Shanghai 201203,China)

机构地区：[1]复旦大学计算机科学与技术系,上海201203

出　　处：《计算机应用与软件》2021年第12期298-303,349,共7页Computer Applications and Software

摘　　要：探索大规模稀疏数据背景下的交叉特征,对提高推荐系统的点击率预测精度十分重要,现有方法往往通过显式暴力枚举或隐式DNN提取完成特征交叉,其掺杂了大量无用、冗余特征,极大地限制了点击率预测模型的表现。对此提出一种显式和隐式相结合的新型特征交叉网络(Deep&Attention Cross Network,DACN)。通过引入基于注意力机制的动态交叉网络,实现以线性空间复杂度完成指定阶显式特征组合,同时消除无用、冗余特征带来的影响。经过实验验证,DACN在预测准确性和参数占用方面相较于现有方法都有提升。Exploring the combinatorial features in the context of large-scale sparse data are essential for improving the accuracy of click through rate prediction in recommendation system.The existing methods usually use explicit violence enumeration or implicit DNN extraction to complete feature combination,which are mixed with a large number of useless and redundant features,and greatly limits the performance of CTR model.This paper proposes a new explicit and implicit network architecture deep&attention cross network(DACN).It introduced a dynamic cross network based on attention mechanism to realize the combination of specified order explicit features with linear space complexity,and eliminated the influence of useless and redundant features.Experimental results show that DACN is better than the existing methods in prediction accuracy and parameter occupancy.

关 键 词：特征组合 注意力机制 特征筛选 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]