RecSys Challenge 2023

Predicting Conversion Rate in Advertising Systems: A Two-Stage Approach with LightGBM
Lulu Wang, Yu Zhang, Huayang Zhao, Zhewei Song, Jiaxin Hu
Paper: https://github.com/colorblank/recsys-challenge-2023/blob/main/recsys2023_challenge.pdf

This project is a solution of Recsys Challenge 2023 provided by the team Ainvest. For more details about this challenge, please visit the official website (https://sharechat.com/recsys2023).

Rank: 4th at Company Leardboard

Environment Setup

Please refer to requirements.txt for environment installation. Note: The best result is obtained using GPU version of LightGBM.

Project Structure

recsys-challenge-2023
├── README.md
├── baseline+2stage.py
├── baseline.py
├── config.py
├── data
│   ├── date_mod_f_11_le
│   ├── f_4_6_le
│   ├── test
│   └── train
├── features.py
└── requirements.txt

Steps to Reproduce

In this challenge, we first utilize LightGBM in conjunction with feature engineering to establish a strong baseline model. Then, we adopt a two-stage modeling approach to estimate the download probability. Finally, we employ an ensemble method to incorporate a deep learning model for enhanced diversity.

Download Dataset

# 1. download from official website
wget https://cdn.sharechat.com/2a161f8e_1679936280892_sc.zip
# 2. unzip file
unzip 2a161f8e_1679936280892_sc.zip
# 3. move files to target folder
mv sharechat_recsys2023_data/test/* data/test/
mv sharechat_recsys2023_data/train/* data/train/

# 4. remove zip and other files
rm -r sharechat_recsys2023_data/
rm -r __MACOSX/
rm 2a161f8e_1679936280892_sc.zip

The data folder:

data/
├── date_mod_f_11_le
├── f_4_6_le
├── test
│   └── 000000000000.csv
└── train
    ├── 000000000000.csv
    ├── 000000000001.csv
    ├── 000000000002.csv
    |__ ...........
    ├── 000000000026.csv
    ├── 000000000027.csv
    ├── 000000000028.csv
    └── 000000000029.csv

Train & Test

1. (Optional) train a baseline model with lightgbm: python baseline.py. The submission result obtain 6.01.

2. train a two-stage model with lightgbm: python baseline+2stage.py. The submission result obtain 5.96.

3. Calibration on result score is_installed: $$ y' = \frac{y}{y + (1- y) / w}, $$ where $w$ set to 0.958.

Final Online Score: 5.949816.