基于深度学习建模多行为动态演化的转化率预估

Conversation rate estimation based on user multiple dynamic behaviors extracted with attention enhanced deep learning

针对基于用户行为特征的转化率预估在计算广告领域的应用中尚未充分提取和利用用户多种行为模式的动态演化特性等问题,考虑用户兴趣模型和行为模型的动态演化性,提出一种融合注意力机制的深度学习策略,获取用户行为动态演化特征,进而构建基于用户行为特征的转化率预估模型.首先,构建基于门控循环单元(gated recurrent unit,GRU)和注意力机制的用户单个行为序列模型,将提取出的用户行为嵌入表示作为用户行为的动态变化特征;然后,利用自注意力对用户的多行为动态演化进行建模;最后,融合所提取的用户多行为序列向量作为用户的行为特征,构建移动APP广告转化率预估模型.实验结果表明所提取的用户行为序列特征可有效改善转化率预估效果.

Conversation rate estimation(CRE)based on user s behaviors has been highly concerned in the computational advertising,and achieved corresponding outstanding researches.The dynamic and evolution of user multiple behaviors,however,have not been sufficiently addressed in CRE.Considering the dynamical and evolutionary features of user behaviors,attention enhanced deep learning is proposed to extract the user behaviors features,and then the CRE model is trained with such features.Attention articulated GRU is first constructed to obtain the single behavior series of a user,and such series are expressed as the dynamical and evolutionary features.Then,the self-attention is used to model the multiple behaviors features.The user multiple behavior series are merged into a vector and used to construct the CRE model.The proposed algorithm is applied to the CRE of Tencent s mobile APP advertisements,and the experimental results demonstrate its advantage in further improving the prediction accuracy.