GIFT: Graph-guIded-Feature-Transfer-Network (CIKM22)

This is the source code for the CIKM 22 paper "GIFT: Graph-guIded Feature Transfer for Cold-Start Video Click-Through Rate Prediction"

Method

Graph Construction for Video Feature Transfer

Physical Linkage

Semantic Linkage

Graph-guIded Feature Transfer (GIFT) network

Results

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.

Getting Start

Requirements

• Python >= 3.6.1
• NumPy >= 1.12.1
• TensorFlow >= 1.4.0

Preprocess dataset

• 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

Results

Model	AUC
DNN	0.7103
GIFT (with dot attention)	0.7175
GIFT (with mlp attention)	0.7199

---

Citation

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}
    }