基于高阶特征交互的点击率预估模型的实现

Implementation of click rate prediction model based on high-order feature interaction

传统的低阶特征模型不能充分利用大数据,从多个维度描述数据和用户。专注于高阶特征提取,结合显式和隐式特征交互的点击率预估模型可以利用好大数据的特点。使用Tensorflow框架搭建包含深度神经网络、因子压缩交互网络和多重特征自交互网络结构的模型,使用淘宝展示广告点击率预估数据集进行训练。模型采用对数损失值和ROC曲线下面积作为评价指标,与原始的LR、FM、Deep&Wide等典型模型进行比较,对数损失值降低了0.04,AUC值提高了0.05左右。

Traditional low-level feature models cannot make full use of big data and describe data and users from multiple dimensions.The click-through rate prediction model that focuses on high-level feature extraction combining the interaction of explicit and implicit features can make good use of the characteristics of big data.The Tensorflow framework was used to build a model including a deep neural network,a compressed interaction network and a multiple self-interaction network structure,and the Ad Display/Click data on Taobao.com were used for training.The logarithmic loss value and the area under the ROC curve were used as evaluation indicators,and compared with the original LR,FM,Deep&Wide and other typical models.The logarithmic loss value is reduced by 0.04,and the AUC value is increased by about 0.05.