Simple Demo of MLflow Project with LightGBM

Adult Data Set from UCI Machine Learning Repository

Since this demo project doesn't focus on machine learning algorithms, so only simple feature engineering steps are implemented. More specifically,

• target: transform '>50K' as 1 and '<=50K' as 0
• fill missing categorical value with 'Missing' and apply label-encoder
• fill missing numerical value with median value from training set
• since the data set is imbalanced, AUC is used as the metric

Gradient Boosting model from LigthGBM are implemented. Check the documentation for details.

• Method 1:

$ git clone https://github.com/JifuZhao/mlflow-demo.git
$ mlflow run mlflow-demo \
  -P train_path=absolute path/mlflow-demo/data/train.csv \
  -P test_path=absolute path/mlflow-demo/data/test.csv \
  -P num_boost_round=1000 -P learning_rate=0.1 -P num_leaves=31 \
  -P max_depth=-1 -P min_data_in_leaf=20

Note: You need to change the file path to be your local absolute path.

After the code successfully runs, you can get the following results:

• Method 2:

$ git clone https://github.com/JifuZhao/mlflow-demo.git
$ cd mlflow-demo/
$ python train.py ./data/train.csv ./data/test.csv 1000 0.1 31 -1 20

After running, you can view the result through MLflow UI:

$ mlflow ui

In the browser, go to http://127.0.0.1:5000, you will get the following results.

• mlflow/mlflow-example
• mserrate/mlflow-sample
• dmatrix/jsd-mlflow-examples

Copyright @ Jifu Zhao 2018