vmipshu / bagfn

This is an official implementation of feature interaction for BaGFN

BaGFN: Broad Attentive Graph Fusion Network for High-Order Feature Interactions

• PDF Link

Introduction

Modeling feature interactions is of crucial significance to high-quality feature engineering on multi-filed sparse data. At present, a series of state-of-the-art methods extract cross features in a rather implicit bit-wise fashion and lack enough comprehensive and flexible competence of learning sophisticated interactions among different feature fields. Broad Attentive Graph Fusion Network (BaGFN) better models high-order feature interactions in a flexible and explicit manner. On the one hand, we design an attentive graph fusion module to strengthen high-order feature representation under graph structure. The graph-based module develops a new bilinear-cross aggregation function to aggregate the graph node information, employs the self-attention mechanism to learn the impact of neighborhood nodes, and updates the high-order representation of features by multi-hop fusion steps. On the other hand, we further construct broad attentive cross module to refine high-order feature interactions at a bit-wise level. The optimized module designs a new broad attention mechanism to dynamically learn the importance weights of cross features and efficiently conduct the sophisticated high-order feature interactions at the granularity of feature dimensions.

Performance

Experiment logs.

Since we repeated experiments and adjusted parameters, the performances are better than those in the paper:

Avazu:

Criteo:

Requirements

The code runs on the following environment:

• Tensorflow 1.14
• python 3.6
• CUDA 10.0+ (For GPU)
• Other requirements: pip install -r requirements.txt

Usage

Data Process:

• Datasets (Avazu and Criteo) can be downloaded from Kaggel Avazu, Kaggle Criteo
• Please refer to data process.

How to run:

• Since the train_examples.txt is tiny, we run the training with batch_size=4: CUDA_VISIBLE_DEVICES=0 python train.py --batch_size 4
• Default training configs can be modified in function parse_args of train.py.
• Feel free to run the code with specifically parameters on whole dataset:

CUDA_VISIBLE_DEVICES=0 python train.py \
                        --embedding_size 20 \
                        --num_gnn 3 \
                        --epoch 3 \
                        --batch_size 1024 \
                        --learning_rate 0.001 \
                        --optimizer_type adam \
                        --field_size 39 \
                        --checkpoint_dir ./checkpoints/Criteo \
                        --log_dir ./logs/Criteo \
                        --data Criteo  \
                        --data_path ./data/Criteo/   \
                        --run_times 1

Docker demo

Have a fast start with docker demo.

Citation

If you find BaGFN useful for your research, please consider citing the following paper:

@ARTICLE{9565146,
  author={Xie, Zhifeng and Zhang, Wenling and Sheng, Bin and Li, Ping and Chen, C. L. Philip},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={BaGFN: Broad Attentive Graph Fusion Network for High-Order Feature Interactions}, 
  year={2023},
  volume={34},
  number={8},
  pages={4499-4513},
  doi={10.1109/TNNLS.2021.3116209}}

Acknowledgement

This code repository gets inspirations from Song's Autoint.