【摘要】 针对点击率预测模型xDeepFM对低阶特征信息利用不充分且训练参数过多问题，提出一种基于多层次特征交互的点击率预测模型。该模型在DNN与嵌入层之间引入二阶特征交互层，充分融合低阶与高阶特征，并通过二阶特征向量对位相乘的方式减少参数量，提升模型的特征组合能力；在输入特征的预处理过程中引入最大支持的维度参数，避免稀疏特征在独热编码映射时出现维度爆炸。结果表明，在Criteo和Avauz两个真实数据集上本文所提模型AUC指标比xDeepFM模型分别提高了0.007 8和0.004 3。更多还原

【Abstract】 The key to CTR prediction lies in the combination of input features. Aiming at xDeepFM, the most effective CTR prediction model using feature interaction, a CTR prediction model based on multi-level feature interaction was proposed to solve the problem of insufficient use of low-order feature information and too many parameters. In this model, a second-order feature interaction layer is introduced between DNN and the embedding layer to fully integrate low-order and high-order features, and the number of parameters is reduced by means of the second-order feature vector multiplication to improve the feature combination ability of the model. The maximum dimension parameters are introduced in the preprocessing process of the input features to avoid the dimension explosion of sparse features in the single thermal coding mapping. The results show that the AUC index of the proposed model is improved by 0.0078 and 0.0043 compared with xDeepFM model on Criteo and Avauz real data sets, respectively.更多还原