FiBiNet分享1

FiBiNET: Combining Feature Importance

and Bilinear feature Interaction for Click-

Through Rate Prediction

何荣炜

Abstract

• Problem : Many current works calculate the feature interactions in a

simple way such as Hadamard product and inner product and they

care less about the importance of features.

• FiBiNET

into consideration, we introduce a Squeeze-and-Excitation network

(SENET) to learn the weights of features dynamically.

• Besides, feature interaction is a key challenge in CTR prediction field

and many related works calculate the feature interactions in a simple

way such as Hadamard product and inner product.

• We propose a new fine-grained way in this paper to calculate the

feature interactions with the bilinear function.

RELATED WORK

• Factorization Machine and Its relevant variants (FM && FFM)

• Deep Learning based CTR Models

explicit way by proposing a novel Compressed Interaction Network (CIN) part.

SENet Module

• The SENET is proved to be successful in image classification tasks

• Sparse Input and Embedding Layer

• SENET Layer