Skip to content
No description, website, or topics provided.
Python Jupyter Notebook
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


Environment options


env name required default description
DATALAKE_CHANNEL_ID True None DataLake channel id
DATALAKE_TRAIN_FILE_ID True None DataLake file id for training
DATALAKE_VAL_FILE_ID False None DataLake file id for validating


env name required default description
INPUT_FIELDS False None Names of features.
e.g. var_1,var_2,var_3
LABEL_FIELD True None Name of label column.


Parameter doc is available on

Try tuning.

  1. num_leaves
  2. min_data_in_leaf
  3. max_depth

More tuning tips are available on

env name required default description
OBJECTIVE True regression Currently regression and binary applications are supported.
Must be one of [regression, regression_l1, huber, fair, poisson, quantile, mape, gamma, tweedie, binary, multiclass, multiclassova].
IS_CLASSIFICATION False True If True, classification, else regression.
BOOSTING True gbdt Must be one of [gbdt, rf]
NUM_ITERATIONS True 100 Number of boosting iterations. constraints: NUM_ITERATIONS >= 0
LEARNING_RATE True 0.05 Shrinkage rate. constraints: LEARNING_RATE > 0.0
NUM_LEAVES True 31 Max number of leaves in one tree. constraints: NUM_LEAVES > 1
TREE_LEARNER True serial Must be one of [serial, feature, data, voting].
NUM_THREADS False 0 Number of threads for LightGBM. 0 means default number of threads in OpenMP.
DEVICE_TYPE True cpu Device for the tree learning, you can use GPU to achieve the faster learning.
Must be one of [cpu, gpu].
SEED False 42 Random seed.
MAX_DEPTH False -1 Limit the max depth for tree model. This is used to deal with over-fitting when #data is small. Tree still grows leaf-wise. <= 0 means no limit.
MIN_DATA_IN_LEAF False 20 Minimal number of data in one leaf. Can be used to deal with over-fitting. constraints: MIN_DATA_IN_LEAF >= 0
MIN_SUM_HESSIAN_IN_LEAF False 1e-3 Minimal sum hessian in one leaf. Like MIN_DATA_IN_LEAF, it can be used to deal with over-fitting. constraints: MIN_SUM_HESSIAN_IN_LEAF >= 0.0.
BAGGING_FRACTION False 1.0 It likes FEATURE_FRACTION, but this will randomly select part of data without resampling. constraints: 0.0 < BAGGING_FRACTION <= 1.0.
POS_BAGGING_FRACTION False 1.0 Used only in binary application. Used for imbalanced binary classification problem, will randomly sample #pos_samples * pos_bagging_fraction positive samples in bagging. Should be used together with NEG_BAGGING_FRACTION.
NEG_BAGGING_FRACTION False 1.0 Used only in binary application. Same as POS_BAGGING_FRACTION.
BAGGING_FREQ False 0 Frequency for bagging. 0 means disable bagging; k means perform bagging at every k iteration.
BAGGING_SEED False 3 Random seed for bagging.
FEATURE_FRACTION False 1.0 LightGBM will randomly select part of features on each iteration if FEATURE_FRACTION smaller than 1.0. For example, if you set it to 0.8, LightGBM will select 80% of features before training each tree. constraints: 0.0 < FEATURE_FRACTION <= 1.0.
EARLY_STOPPING_ROUNDS False 10 Will stop training if one metric of one validation data doesn’t improve in last EARLY_STOPPING_ROUNDS rounds.
VERBOSITY False 1 Controls the level of LightGBM’s verbosity. < 0: Fatal, = 0: Error (Warning), = 1: Info, > 1: Debug
MAX_BIN False 255 Max number of bins that feature values will be bucketed in. Small number of bins may reduce training accuracy but may increase general power (deal with over-fitting). constraints: MAX_BIN > 1
NUM_CLASS False 1 Number of classes. Used only in multi-class classification applications. constraints: NUM_CLASS > 0
METRIC False "" Metric. Support multiple metrics, separated by ,
METRIC_FREQ False 1 Frequency for metric output. constraints: METRIC_FREQ > 0
NFOLD False 5 Number of folds in CV.
VERBOSE_EVAL False None Whether to display the progress. If None, progress will be displayed when np.ndarray is returned. If True, progress will be displayed at every boosting stage. If int, progress will be displayed at every given verbose_eval boosting stage.
STRATIFIED False True Whether to perform stratified sampling.

Run on local

Use requirements-local.txt.

$ pip install -r requirements-local.txt

Set environment variables.

env type description
ABEJA_ORGANIZATION_ID str Your organization ID.

For Jupyter Notebook

Use train_notebook.ipynb.

You can’t perform that action at this time.