FiBiNET and FiBiNET++: Feature Importance and Bilinear Interaction for Click‑Through Rate Prediction

FiBiNET (Feature Importance and Bilinear feature Interaction NETwork) proposes two key innovations for click‑through‑rate (CTR) prediction: a SENet module that dynamically learns the importance of each feature and a bilinear‑interaction layer that improves feature interaction modeling.

In CTR tasks most features are sparse ID features whose embeddings dominate the model parameters; therefore modeling feature importance is crucial for better utilization of these embeddings.

The architecture consists of an embedding layer converting raw features to embeddings E , a SENet layer producing SENet‑like embeddings V , bilinear‑interaction layers that process E and V to obtain vectors p and q , concatenation of p and q , followed by a DNN and a sigmoid output for click probability.

SENet follows three steps—squeeze (mean pooling), excitation (two fully‑connected layers with activation to generate field‑level weights), and re‑weight (scaling the original embeddings by the learned weights).

The bilinear‑interaction layer goes beyond simple inner‑product or Hadamard product by learning a bilinear function with a parameter matrix, enabling richer second‑order interactions while remaining lightweight.

FiBiNET++ addresses the large parameter count of the original model by introducing Bi‑Linear+, which reduces the output size of the bilinear layer, and SENet+, which adds a fourth “fuse” step (bit‑wise addition and layer normalization) to further enhance feature representations.

Feature preprocessing includes one‑hot encoding for categorical fields and embedding‑multiplication for numeric fields, with layer normalization applied to categorical embeddings and batch normalization to numeric embeddings.

The authors note that the two modules are highly reusable and have been applied in production systems such as the dual‑tower recall in Taobao’s recommendation pipeline.

A reference implementation supporting heterogeneous field embedding sizes is provided on GitHub (https://github.com/QunBB/DeepLearning/blob/main/Recommendation/RANK/fibinet.py).