摘要
随着深度学习的发展,近年来CTR预估模型的研究往往基于深度学习使用不同的特征交叉方式来实现CTR预估模型的性能提升.目前最新最有效的研究成果是xDeepFM,它综合了进行隐式和显式的高阶特征交叉方式的子模型.但经实验发现xDeepFM的子模型选择并不完美,而且子模型的组合策略过于简单.对此,本文提出了一种新模型,不仅改进了子模型的选择,而且用注意力机制改进子模型组合方式.为了方便,在本文中将提出的新模型叫做Attentional-xDeepFM-C.在Avazu和Criteo数据集上进行实验,新模型在两组数据集下的AUC得分分别比xDeepFM模型高2.17%和4.97%.本文已在公开网站#上发布了Attentional-xDeepFM-C模型的源代码.
With the development of deep learning,in recent years,the research of CTR prediction model is often based on deep learning and uses different feature intersection methods to achieve the performance improvement of CTR prediction model. The latest and most effective research result is xDeepFM,which integrates sub-models for implicit and explicit high-order feature intersection methods.However,our experiments have found that the sub-model selection of xDeepFM is not perfect,and the combination strategy of submodels is too simple. In this regard,we propose a new model that not only improves the selection of sub-models,but also uses attention mechanism to improve the combination of sub-models. For convenience,the new model we propose is called Attentional-xDeepFM-C in this article. We conducted experiments on the Avazu and Criteo datasets. The AUC scores of the new model under the two datasets were 2. 17% and 4. 97% higher than the xDeepFM model,respectively. We have released the source code of the AttentionalxDeepFM-C model on open site.
作者
王越
于莲芝
WANG Yue;YU Lian-zhi(School of Optoelectronic Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2021年第9期1884-1890,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61603257)资助。
关键词
因子分解机
CTR预估
推荐系统
注意力机制
深度神经网络
factorization machine
CTR prediction
recommendation system
attention mechanism
deep neural network
