面向点击通过率预测的交互边选择算法研究

Research on interactive edge selection algorithm for CTR prediction

点击通过率(click-through rate,CTR)作为推荐系统中必不可少的核心任务分支,提高其预测准确性,既能改善用户的浏览体验,也能为平台增加收益。以往模型在对点击通过率进行建模预测时,保留所有的交互特征存在信息冗余,交互低效等问题。针对这一问题提出了一种面向点击通过率预测的交互边选择模型,通过自动识别冗余信息来动态选择有益的交互特征,主要由交互边选择网络层,图节点相似度注意力层构成。交互边选择网络层引入过滤阈值机制并结合动态关联矩阵来去除冗余信息,图节点相似度注意力层通过学习相似度权重矩阵来解决节点过度平滑问题。在Criteo和Avazu两个公开数据集上的大量实验证明,该模型的预测能力优于已有模型。

Click-through rate(CTR)is an indispensable core task branch in the recommendation system.Improving its prediction accuracy can not only improve the user’s browsing experience,but also increase the revenue of the platform.Retaining all interactive features brings the problems of information redundancy and inefficient interactions when modeling and predicting CTR in previous models.This paper proposes an interactive edge selection model for CTR,which is used to dynamically select beneficial interactive features by automatically identifying redundant information.It is mainly composed of the interactive edge selection layer and the graph node similarity attention layer.The redundant information is removed by introducing a filtering threshold mechanism combined with a dynamic correlation matrix in interactive edge selection layer.Then the problem of excessive smoothing of nodes is solved by learning the similarity weight matrix in graph node similarity attention layer.Lastly,a large scale of experiments on the Criteo and Avazu datasets demonstrate that the proposed model has better predictive ability than the existing models.