Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)

时间:2019.05

作者:Tongwen Huang, Zhiqi Zhang, Junlin Zhang

 

Abstract

文中使用Squeeze-Excitation Network(SENET)动态学习特征的重要性,并且使用bilinear function学习特征组合

 

Model Structure

Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)        

SENET Layer

Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)

  • Squeeze:用max/average pooling学习每一维特征的全局信息
  • Excitation:用2个全连接层学习权重                                          Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)
  • Re-weight:                                                                                        Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)

 

Bilinear Interaction

Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)

  • Field-all type:所有特征组合共享一个参数矩阵Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)                                                                                  Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)
  • Field-each type:每个特征组i维护一个参数矩阵Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)                                                                                  Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)
  • Field-interaction type:每对特征组合ij维护一个参数矩阵Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)                                                                                  Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)

Combination Layer

Combining Feature Importance and Bilinear feature Interaction for CTR Prediction (FiBiNET)

 

Reference

https://zhuanlan.zhihu.com/p/72931811