recsys-challenge-2023

Dataset Schema

├── configs
│   ├── default_config.py
│   └── submit.py
├── data
│   └── common.py
├── imgs
├── models
│   └── common.py
├── README.md
├── preprocess.py
├── requirements.txt
└── run.py

Preprocess

python preprocess.py --config submit

Format all csv files to one parquet file.

• Input
  • train 000000000000.csv ~ 000000000030.csv
  • test 000000000000.csv
• Output
  • train.paruqet
  • test.parquet

Train & Inference

python run.py --config submit --no_wandb --verbose

• Input
  • train.parquet
  • test.parquet
• Output
  • submission file
  • config.py
  • feature_importance.png

Solution Summary

• Preprocessing

  • count preprocessing : Divide into the second smallest value.
  • remove features : Using Adversarial Validation (Remove Train-Test gap features)

• Feature Engineering

  • featv1 : Encodes the category for the current month using the number of value counts
    • feat_frequency_encoding_7days : using previous 7 days
    • feat_frequency_encoding_full_daysv3 : using all train data (don't contain test period)
  • featv2 : Encodes the group category for the current month using the number of value counts from the previous month.
    • feat_f_2_4, feat_f_4_6, feat_f_3_15, feat_f_13_15
  • featv3 : Day of week feature
    • feat_week
  • featv4 : f_42 / f_42 of previous period
    • daily_f_42 : using previous all days
    • hyeon_daily_f_42_v1 : using previous 7 days
  • featv5 : Catboost encoding of categorical features using previous all days
    • hyeon_click_cat_encoding : using target click
    • hyeon_install_cat_encoding : using target install
  • Others : f_51 / (f_56 + 1)

• Validation

  • Train : f_1 < 66
  • Valid : f_1 = 66

• Modeling

  • LightGBM with categorical features