Solutions for RecSys Challenge 2023

Team Info

Team name: BUAA_BIGSCity

Team member: Jiawei Jiang, Bing Wang.

Team adviser: Jingyuan Wang.

Rank: 6th in the academic track

Train & Test

You can reproduce the results by following these steps:

Step 1: Download the dataset and place it in the ./data/ directory to get the following directory structure:

./data/test/*.csv
./data/train/*.csv

Step 2: Environment Configuration:

conda create -n recsys python=3.9.7
pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 torchaudio==0.10.1 -f https://download.pytorch.org/whl/cu111/torch_stable.html
pip install -r requirements.txt

Step 3: Fuse the training data set:

python merge_data.py

This command will generate the file ``./data/train.csv`.

Step 4: Run command to train three models:

python trainDeep.py --model xDeepFM --dense_bins 5 --sparse_dim 4 --batch_size 256 --epoch 50 --patience 10 --device 2 --exp_id 96499
python trainDeep.py --model xDeepFM --dense_bins 10 --sparse_dim 4 --batch_size 256 --epoch 50 --patience 10 --device 1 --exp_id 3957
python trainDeep.py --model xDeepFM --dense_bins 5 --sparse_dim 4 --batch_size 256 --epoch 50 --patience 10 --device 1 --dropout 0.2 --exp_id 1984

Each command generates one model, where the exp_id in the command represents the ID record of one training.

Correspondingly, the trained models are stored in ./ckpt/exp_id_*.ckpt, and the model prediction results are stored in ./output/exp_id_*.csv, and the logs of the training process are stored in ./log/exp_id_*.log.

Step 5: Merge the output of 3 models. [This step can be skipped because the predicted results are already included in the ./output directory.]

python result_merge.py

This command will fuse the output of the 3 models to get the file avg_96499_3957_1984_1_1_2.csv. This file has a final submission score of 6.282142, ranking 6th in the academic track.

Method Introduction

The proposal will be presented in the paper submitted to the workshop.

Acknowledgements

DeepCTR-Torch