recsys-challenge-2024

This is the code implementation of RecSys Challenge 2024 focusing on online news recommendation by Ekstra Bladet and JP/Politikens Hus A/S ("Ekstra Bladet").

Our organization is hrec (captain: doubleQ).

Usage

Machine: 36 cores, 2T memory, 2 * V100

Unzip the data files in ./data/

data/
├── image_embeddings.parquet
├── contrastive_vector.parquet
├── small
│   ├── train
│   ├── validation
│   ├── articles.parquet
├── large
│   ├── train
│   ├── validation
│   ├── articles.parquet
├── ebnerd_testset
│   ├── test
│   ├── articles.parquet

Generate base features

• python preprocess.py
• The base features will be in preprocessed/large

Generate din and dcn prediction scores based on the preprocessed features

• cd fuxictr
• python convert_data.py # generate fuxictr features in Ebnerd_large_data/
• python run_expid.py --config config/ebnerd_large_tuner_config --expid DCN_Ebnerd_large_001 --gpu 0
• python run_expid.py --config config/ebnerd_large_tuner_config --expid DIN_Ebnerd_large_001 --gpu 0 This step will generate scores in fuxictr/features

Train xgboost/lightgbm with pairwise loss or BCE base on these features

• python main.py --model_name xgb --mode large --rank will get result in result/xgb_submit.zip
• The final result will generated within result/xgb_submit.zip
• It can achieve AUC 0.8714 in the official leaderboard, which was however failed to be submitted to the competition leaderboard due to the Insufficient Space issue, because I deleted it for more space.