1.Embedding and Stacking Layer(嵌入和堆叠层)

离散特征embedding:将二进制的离散特征通过embedding转换成实数值的稠密向量,稠密特征归一化

嵌入向量与归一化稠密特征叠加起来形成一个向量作为输入

http://upload-images.jianshu.io/upload_images/14619338-adfe28a1fe11e377.png?imageMogr2/auto-orient/strip|imageView2/2/w/1240

2.Cross Network Layer (交叉网络层)

Embedding and Stacking Layer作为输入,执行

https://upload-images.jianshu.io/upload_images/14619338-ec9e9cb8f285589c.png?imageMogr2/auto-orient/strip|imageView2/2/w/1240

3.Deep Network Layer (深度网络层)

全连接的前馈神经网络:

https://upload-images.jianshu.io/upload_images/14619338-b7f6a745fab64f35.png?imageMogr2/auto-orient/strip|imageView2/2/w/1240

4.Combination Layer (连接层)

将 Cross Network Layer和Deep Network Layer输出的向量连接,输入到逻辑回归模型sigmoid()中

https://upload-images.jianshu.io/upload_images/14619338-a2d5210cc6306bd7.png?imageMogr2/auto-orient/strip|imageView2/2/w/1240