Barebones machine learning pipeline

Before running locally, start a Postgres database locally or remotely and add the host/domain/user/password information to config_db.yaml. Within the same file you can change the input_path to point to data_full.csv insteal of data_small.csv. The smaller file contains 5,000 rows and is much faster for testing in a local environment.

git clone https://github.com/timhannifan/minml
cd minml
sh run.sh

To run in ipython:

virtualenv env
source env/bin/activate
pip install -r requirements.txt
cd src/minml

ipython3
run main --config "../../examples/donors/config/config_test_env.yaml" --db "../../examples/donors/config/config_db.yaml"

In the main config file there is a load_db parameter that prevents table generation and re-reading csv data. After the first run, run without this set to False to improve performance. In addition, featurized datasets are saved to the project directory using a toggle in the main config, which allows you to avoid re-generating features on a dataset.

See requirements.txt for required packages. The current stack requires:

• Python 3.6.0
• Postgres 11.3

Minml Experiments are configured via a set of YAML and SQL files within a project directory. See examples for test/production environments .

The project directory should contain all the files listed in the example directory, with the same filenames. The main configuration file is config.yaml, which defines the Experiment parameters for creating time splits, generating features, specifying models and parameter ranges, and selecting a scoring matrix for model evaluation.

To run a custom experiment with new data, one would need to replace the code within each file prefixed with config_ in the project directory. The SQL files define the Experiment schemas, and field-specific information for indexing, cleaning, inserting, and defining semantic entities/events.

Results from model/parameter/threshold/time-split analysis are stored in a Postgres table named 'results'. A sample of the results_exploration.sql output is shown below.

Precision/recall graphs vs population thresholds for the best performing models in each split are exported to a directory specified in the config. Visualizations can also be disabled in the config.

Work is underway for post-modeling evaluation and parameter tuning.