基于改进Fi-GNN模型的点击率预测方法
Click rate prediction method based on improved Fi-GNN model
夏义春 1李汪根 1李豆豆 1高坤 1束阳1
作者信息
- 1. 安徽师范大学计算机与信息学院,安徽 芜湖 241002
- 折叠
摘要
为解决基线模型(Fi-GNN)特征交互模块设计不合理的问题,提出一种基于改进Fi-GNN模型的点击率预测方法(Fi-GNN-V2).针对特征交互模块的邻接矩阵没有考虑到异构节点间的多元关系,在计算异构节点间相互作用的权重时增加边类型的嵌入向量,得到更合理的邻接矩阵;通过多头聚合多个子空间的邻居信息学习不同方式的特征交互;融合二阶以及三阶特征组合解决特征交互模块造成特征域的语义信息丢失问题,设计注意力模块抑制无用特征组合对模型学习的干扰;为进一步提升模型的性能,结合深度神经网络隐式捕捉高阶非线性的特征组合进行联合预测.实验结果表明,该方法优于其它主流点击率预测模型.
Abstract
To solve the problem of unreasonable design of feature interaction module of baseline model(Fi-GNN),a click through rate prediction method based on improved Fi-GNN model(Fi-GNN-V2)was proposed.The adjacency matrix of feature interac-tion module does not take the multiple relationships between heterogeneous nodes into account.Therefore,when calculating the interaction weight between heterogeneous nodes,the embedded vector of edge types was added to obtain a more reasonable adjacency matrix.Different ways of feature interaction were learned by aggregating neighbor information of multiple subspaces.The combination of second-order and third-order features was fused to solve the problem of semantic information loss in feature domain caused by feature interaction module,and the attention module was designed to suppress the interference of useless fea-ture combination on model learning.To further improve the performance of the model,the model was combined with deep neural network to implicitly capture high-order nonlinear feature combinations for joint prediction.Through the comparative experi-ments between the two benchmark data sets and the baseline model and several advanced mainstream models,experimental results show that this method is superior to other mainstream click-through prediction models.
关键词
点击率预测/邻接矩阵/异构节点/多空间聚合/语义信息/注意力模块/深度神经网络Key words
click through rate prediction/adjacency matrix/heterogeneous node/multi spatial aggregation/semantic informa-tion/attention module/deep neural network引用本文复制引用
基金项目
国家自然科学基金(61976006)
高校领军人才引进与培育计划(051619)
出版年
2024
NETL
NSTL
万方数据