Tackling Covariate Shift in ML Using ML

Published:

In the previous post I mentioned about a simple way of estimating the density ratio of two probability distributions. I decided to create a python package that provides such a functionality.

You can find the repo here.

The README should have explained the details already. However, let me post it again in this post.

Modules

The important modules:

• probabilistic_classification_covshift - the main module
• automl.trainer - find the best classifier that separates the source instances from the target ones
• automl.predictor - the best classifier is used to compute the probability of each instance belongs to the source or target origin. The computed probabilities become the parameters for the weight calculation

What you need

• python
• pyspark
• h2o and its dependencies:
• requests
• tabulate
• "colorama>=0.3.8"
• future

Install

1. Clone this repo
2. Go to the repo’s root directory
3. Run the following command: python setup.py install

Quickstart

You might want to take a look at the example.

A) Compute weight

Prepare the configuration for AutoML.

conf = {
AutoMLConfig.DATA: {
AutoMLConfig.LABEL_COL: 'label',
AutoMLConfig.ORIGIN_COL: OriginFeatures.ORIGIN,
AutoMLConfig.WEIGHT_COL: WeightFeatures.WEIGHT,
AutoMLConfig.BASE_TABLE_PATH: 'data/base_table.parquet',
AutoMLConfig.WEIGHT_PATH: 'data/weight.csv'
},
AutoMLConfig.SERVER_CONN_INFO: {
H2OServerInfo.IP: 'localhost',
H2OServerInfo.PORT: '54321'
},
AutoMLConfig.CROSS_VAL: {
AutoMLConfig.FOLD_COL: "fold",
AutoMLConfig.NFOLDS: 8,
},
AutoMLConfig.MODELING: {
AutoMLConfig.MAX_RUNTIME_SECS: 3600,
AutoMLConfig.MAX_MODELS: 10,
AutoMLConfig.STOPPING_METRIC: 'logloss',
AutoMLConfig.SORT_METRIC: 'logloss'
},
AutoMLConfig.EXCLUDE_ALGOS: [
"StackedEnsemble",
"DeepLearning"
],
AutoMLConfig.MODEL: {
AutoMLConfig.BEST_MODEL_PATH: 'data/model/'
},
AutoMLConfig.SEED: 23
}

Run the probabilistic classification module.

source_df = <spark_dataframe>
target_df = <spark_dataframe>

pc = ProbabilisticClassification(source_df, target_df, conf)
pc.run()

B) Append the weights to the base table

We got the weights! They are stored as a csv file in a location specified by conf[AutoMLConfig.DATA][AutoMLConfig.WEIGHT_PATH].

Now, we just need to append them to the base table. The base table could be the source data, target data, or merged data (source and target). Please adjust with your needs.

Suppose that we’d like to append the weights to the merged data.

base_table_path = conf[AutoMLConfig.DATA][AutoMLConfig.BASE_TABLE_PATH]
weight_path = conf[AutoMLConfig.DATA][AutoMLConfig.WEIGHT_PATH]
origin_col = conf[AutoMLConfig.DATA][AutoMLConfig.ORIGIN_COL]

weighted_base_frame_df = base_frame_df.join(weight_df, how='left', on='row_id').drop('row_id')

How if we’d like to append the weights to the source data only?

source_df = base_frame_df.filter(F.col(origin_col) == OriginFeatures.SOURCE)

weighted_base_frame_df = source_df.join(weight_df, how='left', on='row_id').drop('row_id')

Done.

Contribute

All features requests, documentations or bugs fixes for future improvement are welcomed.

Simply do the followings:

• Fork this repo
• Create a local branch
• Develop your features on the branch
• Submit a pull request