This is the source code for the CIKM 22 paper "GIFT: Graph-guIded Feature Transfer for Cold-Start Video Click-Through Rate Prediction"
We report the results on a large scale video recommendation dataset collected from Guess You Like (猜你喜欢) of Taobao App's homepage. Online evaluation shows that GIFT has brought 6.82% lift on CTR metric (from 4.180% to 4.465% during Sep.21 ~ Sep.27, 2020).
Model | AUC |
---|---|
DNN | 0.7423 |
DeepFM | 0.7508 |
DIN | 0.7568 |
GIFT | 0.7670 |
GIFT with finetune | 0.7693 |
Due to Alibaba Group's privacy policy, we cannot publish the source dataset used in the paper, but we re-implement the GIFT network and conduct experiments on the DBook dataset, which is collected from www.douban.com.
- Python >= 3.6.1
- NumPy >= 1.12.1
- TensorFlow >= 1.4.0
- Step 1: Construct physical linkages between new books and old books.
cd FeatGeneration;
python DBook_graph_construction.py;
python DBook_data_process.py;
or
sh process_dbook.sh;
- Step 2: Train a GIFT (based on DNN) model
cd Train;
python train_dnn_gift.py;
python eval_dnn_gift.py;
- Step 3: Train and evaluate a DNN model
cd Train;
python train_dnn.py;
python eval_dnn.py
Model | AUC |
---|---|
DNN | 0.7103 |
GIFT (with dot attention) | 0.7175 |
GIFT (with mlp attention) | 0.7199 |
Slides
Here is our slides: https://github.com/Bayi-Hu/GIFT-Graph-guided-Feature-Transfer-Network/blob/master/materials/GIFT_CIKM22_slides.pdf
Bibtex:
@article{hu2022gift,
title={GIFT: Graph-guIded Feature Transfer for Cold-Start Video Click-Through Rate Prediction},
author={Hu, Sihao and Cao, Yi and Gong, Yu and Li, Zhao and Yang, Yazheng and Liu, Qingwen and Ou, Wengwu and Ji, Shouling},
journal={arXiv preprint arXiv:2202.11525},
year={2022}
}