摘要
在推荐系统研发中,点击率(Click-Through Rate,CTR)预估是非常重要的工作,点击率预估精度的提升直接影响到整个推荐系统的收益,对其性能和解释性的研究有助于理解系统决策的机理,同时还能帮助优化需求和系统设计。当前点击率预估深度模型多基于线性特征交互和深度特征提取进行设计。由于深度模型的黑盒特点,该类模型在解释性方面存在局限性,并且在先前的研究中,对点击率预估模型的解释性研究非常少。因此,文中基于多头自注意力机制,对该类模型的解释性进行研究,通过多头注意力机制对特征嵌入、线性特征交互和深度部分进行增强和解释,在深度部分设计了两种模型,即注意力增强的深度神经网络和注意力叠加的深度模型,通过计算每个模块的注意力得分对其进行解释。所提方法在多个真实数据集上进行了大量实验,结果表明所提方法能够有效提升模型效果,并且模型自身带有一定的解释性。
Click-Through Rate(CTR)prediction is critical to recommender systems.The improvement of CTR prediction can directly affect the earnings target of the recommender system.The performance and interpretation of the CTR prediction algorithm can guide developers to understand and evaluate recommender system accurately.That's also helpful for system design.Most existing approaches are based on linear feature interaction and deep feature extraction,which have poor model interpretation in the outcomes.Moreover,very few previous studies were conducted on the model interpretation of the CTR prediction.Therefore,in this paper,we propose a novel model which introduces multi-head self-attention mechanism to the embedding layer,the linear feature interaction component and the deep component,to study the model interpretation.We propose two models for the deep component.One is deep neural networks(DNN)enhanced by multi-head self-attention mechanism,the other computes high-order feature interaction by stacking multiple attention blocks.Furthermore,we calculate attention scores and interpret the prediction results for each component.We conduct extensive experiments using three real-world benchmark datasets.The results show that the proposed approach not only improves the effect of DeepFM effectively but also offers good model interpretation.
作者
杨斌
梁婧
周佳薇
赵梦赐
YANG Bin;LIANG Jing;ZHOU Jiawei;ZHAO Mengci(China Unicom Research Institute,Beijing 100048,China;School of Computer Science,Beijing University of Posts and Telecommunications,Beijing 100876,China;School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876,China)
出处
《计算机科学》
CSCD
北大核心
2023年第5期12-20,共9页
Computer Science
关键词
推荐系统
点击率预估
多头自注意力机制
特征交互
模型解释性
Recommender system
Click-Through Rate prediction
Multi-head self-attention mechanism
Feature interaction
Model interpretability
