Code for SATrans

The source code for our paper "Scenario-Adaptive Feature Interaction for Click-Through Rate Prediction" (accepted by KDD2023 Applied Science Track). The preprocessed dataset can be downloaded from link.

--data_name: name of dataset
--model_name: name of model
--domain_col: the name of the feature indicating the domain, e.g., in aliccp dataset, feature column "301" is the domain feature
--domain_att_layer_num: Transformer layer number
--att_head_num:  the number of heads in self-attention layers
--flag:  the flag of different functions
--meta_mode:  apply meta_mlp over Q/K/V, should be a string,e.g., "QK","Q","QKV",where "QK" has the best performance 

• Run SATrans(EN+MetaNet) on aliccp dataset (Best performance)

python main.py --data_name alicpp --model_name SATrans --seed 1021 --embedding_dim 32 --learning_rate 0.005 --domain_att_layer_num 3 --att_head_num 4 --meta_mode QK --domain_col 301 --flag sota

• Run SATrans(ENP+MetaNet) on alimama dataset (use --flag to set "pos") (Best performance)

python main.py --data_name alimama --model_name SATrans --seed 1021 --embedding_dim 32 --learning_rate 0.001 --domain_att_layer_num 3 --att_head_num 4 --meta_mode QK --domain_col shopping_level --flag sota-pos

Environments

• NVIDIA GeForce GTX 1080 Ti
• CUDA 11.2 (For GPU)
• Python==3.7.10
• torch==1.10.0+cu113
• tensorflow==2.7.0
• deepctr-torch==0.2.9
• scikit-learn==1.0.1

Reference

Please cite the paper whenever this code is used to produce published results or incorporated into other software:

@inproceedings{min2023scenario,
  title={Scenario-Adaptive Feature Interaction for Click-Through Rate Prediction},
  author={Min, Erxue and Luo, Da and Lin, Kangyi and Huang, Chunzhen and Liu, Yang},
  booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={4661--4672},
  year={2023}
}