FiBiNet++:Improving FiBiNet by Effectively Reducing Model Parameters for CTR Prediction

Code of FiBiNet++:Improving FiBiNet by Effectively Reducing Model Parameters for CTR Prediction.

Also support FiBiNet of paper FiBiNET: Combining Feature Importance and Bilinear feature Interaction for Click-Through Rate Prediction .

Prerequisites

• Python == 3.6
• TensorFlow-GPU == 1.14

Getting Started

Installation

• Install TensorFlow-GPU 1.14

• Clone this repo

Dataset

• Links of datasets are:

  • https://www.kaggle.com/c/criteo-display-ad-challenge
  • https://www.kaggle.com/c/avazu-ctr-prediction

• You can download the original datasets and preprocess them by yourself. Run python -u -m fibinet.preprocessing.{dataset_name}.{dataset_name}_process to preprocess the datasets. dataset_name can be criteo or avazu.

• This repo also contains a demo dataset of criteo, which contains 100,000 samples and has been preprocessed. It is used to help demonstrate FiBiNet++.

Training

You can use python -u -m fibinet.run --version {version} --config {config_path} to train a specific model on a dataset.

Some important parameters are list below, and other hyper-parameters can be found in the code.

• version: model version, supports v1, ++, and custom, default to ++. For custom, you can adjust all parameter values flexibly.
• config_path: specifies the paths of the input/output files and the fields of the dataset. It is generated during dataset preprocessing. Support values: ./config/criteo/config_dense.json, ./config/avazu/config_sparse.json.
• mode: running mode, supports train, retrain, test.

Acknowledgement

Part of the code comes from DeepCTR.