My Solution to atmaCup #8 (private 29th, public 31st)

Competition Page: atmaCup #8

Solution Summary

• Models: Linear Stacking of GBDTs (LGB, CatB, XGB)

• Validation Strategy: StratifiedKFold, using the binned target (Global_Sales)

• Features: ~1,500 features in total, mainly:

  • PCA, NMF, LDA for co-occurence matrices for categorical features
  • word2vec for categorical features (Category2VecWithW2V)
  • tf-idf to pca for Name (using texthero)
  • aggregation (mean, min, max etc) for numerical features (Critic_score etc) by categorical features (Platform etc)
  • further statistics computed from the abovementioned aggregation features (z-score, max to min ratio etc)
  • nunique by categorical features for the other categorical features

It turned out that EDA (Exploratory Data Analysis) is the key to win this competition, as is always the case with a tabular data science competition!

Usage

Folders

Confirm that the following folders are located in the exact manner. Add the competition data into the input folder.

.
├── script
├── config
├── output
└── input

How to Use

I use my local Mac, employing

• docker
• mlflow
• hydra

to manage my experiments.

Run the followings in the docker container (which is mentioned later) to get a submission csv.

python script/feature_engineering.py
python script/fit_predict.py

Environment

Docker

Install docker / docker-compose in your machine.

(base) docker-compose up -d --build  # build a container
(base) docker exec -it katsu1110-atmacup8 bash # get in a container

• By typing jupyter lab inside the container jupyter lab can be initiated in http://localhost:8888/.
• By typing mlflow ui inside the container mlflow ui can be initiated in http://localhost:5000/.