LPM.discretize

Overview

A library and CLI tool for taking CSV files producing new CSV files where all numerical columns have been discretized.

Disclaimer

This is pre-alpha software. We are currently testing it in real-world scenarios. In its present state, we discourage users from trying it.

Installation

This library is packaged with Nix Flakes. Include this as a Python module in your Flakes file or use the CLI directly (see examples below).

Usage

Users can discretize a single dataframe or a list of dataframes.

If multiple dataframes are supplied, one will be used as a reference point. This is necessary for example when quantile based discretization is applied.

They can supply custom functions for discretization or use quantile-based discretization.

Using LPM.discretize as a Python library

Discretize a single dataframe based on quantiles

Tak a single dataframe and discretize numerical columns based on a quantiles:

# Get dependencies.
import polars as pl
from lpm_discretize import discretize_df_quantiles

# Read a csv files.
df = pl.read_csv("data.csv")

# Discretize this dataframe; all columns are discretized based on 4 quantiles. 
df_discretized =  discretize_df_quantiles(df, quantiles=4)

Users can list which columns will be discretized. Let's discretize only foo and bar.

df_discretized =  discretize_df_quantiles(df, quantiles=4, columns=["foo", "bar"])

Note that the quantiles argument here is overloaded. Users can supply an int to set the number of quantiles for every column, or they supply a dictionary.

df_discretized =  discretize_df_quantiles(df, quantiles={"foo": 4, "bar": 2}))

Discretize a single dataframe based on a map columns->discretization-functions

Users can supply their own discretization functions as dictionaries.

from lpm_discretize import discretize_df
discretize(df, discretization_functions={"foo": lambda x: "yes" if x <= 1 else "no", "bar": lambda x: "ja" if x <= 6 else "nein"})

Of course, this can be wrapped into a list comprehension discretizing multiple dataframes.

Discretizing multiple dataframes based on quantiles

import polars as pl
from lpm_discretize import discretize_df_quantiles

# Read three csv files.
df_a = pl.read_csv("real-data.csv")
df_b = pl.read_csv("synthetic-data-lpm.csv")
df_c = pl.read_csv("synthetic-data-baseline.csv")

# Get discretized version of the dataframes above. Use Polar's types do decide what to discretize.
# By default, this discretizes with based on quartiles in `df_a`.
df_a_discretized, df_b_discretized, df_c_discretized =  discretize_quantiles([df_a, df_b, df_c], quantiles=4)

Below use a list of columns to discretize
df_a_discretized, df_b_discretized, df_c_discretized =  discretize_quantiles(
    [df_a, df_b, df_c],
    quantiles=4,
    columns=["foo", "bar"]
)

See docstrings for other usage patterns.

LPM.discretize CLI

Usage information for the CLI can be printed with the following command.

nix run github:InferenceQL/lpm.discretize --help

Usage information for the CLI can be printed with the following command, this internally invokes discretize_df_quantiles.

nix run github:InferenceQL/lpm.discretize -- --data data.csv

For customization purposes, we expect users to use this tool as a Python library. See above.

Test

Tests are automatically run through the flakes file.

During development, uses can either add Pytest to the flakes output and use the Nix shell:

nix develop -c  pytest tests/ -vvv

or they install the library globally.

python -m pip install --upgrade --force-reinstall  . && pytest tests/ -vvv

The latter worflow depends on pip and pytest being available globablly, too.