Laboratory of Machine Learning

This command-line app is designed to perform a battery of regression tests on a specific research dataset. It is not a general-purpose library, but rather a case-study or template for a specific type of analysis.

The application has two main commands:

• prepare-data: This command generates and saves CSV files containing integer indices for cross-validation folds. It does not save preprocessed data. The logic is tightly coupled to a specific dataset structure, expecting an Excel file where features start at column 6, and the target variable is in column 5.
• regression: This command runs an exhaustive grid search over a hardcoded set of scalers, dimensionality reduction techniques, and regression models. It can either generate cross-validation splits on the fly or use the index files created by the prepare-data command.

The result is a general idea of how well these techniques work with different standardization techniques and some dimension reduction on the different machine learning approaches.

After that, the researcher is encouraged to explore a fine-tuning of the more prominent techniques to obtain the best model.

There is also an unregistered, standalone script get_results_partition.py for running a manually-defined list of specific pipelines on a single train-test split.

TODO list:

• Include a RandomSearch in the process to select also the more prominent approaches
• Add classification techniques
• Make the split and treatment of the problem agnostic of the file.
• Allow the inclusion of other approaches in a more organic way (for example, a config file)

Dependencies (python)

• pandas
• numpy
• scikit-learn
• pathlib
• rich
• typer
• scipy
• openpyxl

Installation and Usage

1. Clone the repository.

2. Install the dependencies and the command-line tool:

   uv sync

   [!NOTE] After modifying pyproject.toml or setup.py, you may need to run this command again to make new scripts available.

3. Run the application:

   uv run labml --help

   For example, to run the regression command:

   uv run labml regression --help