Autoimpute

Autoimpute is a Python package for analysis and implementation of Imputation Methods!

View our website to explore Autoimpute in more detail. New tutorials coming soon!

Check out our docs to get the developer guide to Autoimpute.

Conference Talks

We presented Autoimpute at a couple of PyData conferences!
PyData NYC: New and Upcoming slot in November 2019
PyData LA: Main talk slot in December 2019. See the video here!

Notes on Development

v 0.13+ has upgraded to support pymc, the next generation of the pymc3 library
v 0.12+ contains the new MiceImputer. Thanks to @gjdv for the help on this issue!
We're looking to collaborate and happy to work with those interested!
If you'd like to get involved, feel free to reach out! Our info is on the Authors page.

Creators and Maintainers

Joseph Kearney – @kearnz
Shahid Barkat - @shabarka
Arnab Bose (Advisor) - @bosearnab
See the Authors page to get in touch!

Installation

Autoimpute is registered with PyPI! Download with pip install autoimpute.
If pip cached an older version, try pip install --no-cache-dir --upgrade autoimpute.
If you want to work with the development branch, use the script below:

Development

git clone -b dev --single-branch https://github.com/kearnz/autoimpute.git
cd autoimpute
python setup.py install

Motivation

Most machine learning algorithms expect clean and complete datasets, but real-world data is messy and missing. Unfortunately, handling missing data is quite complex, so programming languages generally punt this responsibility to the end user. By default, R drops all records with missing data - a method that is easy to implement but often problematic in practice. For richer imputation strategies, R has multiple packages to deal with missing data (MICE, Amelia, TSImpute, etc.). Python users are not as fortunate. Python's scikit-learn throws a runtime error when an end user deploys models on datasets with missing records, and few third-party packages exist to handle imputation end-to-end.

Therefore, this package aids the Python user by providing more clarity to the imputation process, making imputation methods more accessible, and measuring the impact imputation methods have in supervised regression and classification. In doing so, this package brings missing data imputation methods to the Python world and makes them work nicely in Python machine learning projects (and specifically ones that utilize scikit-learn). Lastly, this package provides its own implementation of supervised machine learning methods that extend both scikit-learn and statsmodels to mutiply imputed datasets.

Main Features

Utility functions to examine patterns in missing data and decide on relevant features for imputation
Missingness classifier and automatic missing data test set generator
Native handling for categorical variables (as predictors and targets of imputation)
Single and multiple imputation classes for pandas DataFrames
Custom visualization support for utility functions and imputation methods
Analysis methods and pooled parameter inference using multiply imputed datasets
Numerous imputation methods, as specified in the table below:

Imputation Methods Supported

| Univariate | Multivariate | Time Series / Interpolation | :-------------------------- | :---------------------------------- | --------------------------- | Mean | Linear Regression | Linear | Median | Binomial Logistic Regression | Quadratic | Mode | Multinomial Logistic Regression | Cubic | Random | Stochastic Regression | Polynomial | Norm | Bayesian Linear Regression | Spline | Categorical | Bayesian Binary Logistic Regression | Time-weighted | | Predictive Mean Matching | Next Obs Carried Backward | | Local Residual Draws | Last Obs Carried Forward

Todo

Additional cross-sectional methods, including random forest, KNN, EM, and maximum likelihood
Additional time-series methods, including EWMA, ARIMA, Kalman filters, and state-space models
Extended support for visualization of missing data patterns, imputation methods, and analysis models
Additional support for analysis metrics and analyis models after multiple imputation
Multiprocessing and GPU support for larger datasets, as well as integration with dask DataFrames

Example Usage

Autoimpute is designed to be user friendly and flexible. When performing imputation, Autoimpute fits directly into scikit-learn machine learning projects. Imputers inherit from sklearn's BaseEstimator and TransformerMixin and implement fit and transform methods, making them valid Transformers in an sklearn pipeline.

Right now, there are three Imputer classes we'll work with:

from autoimpute.imputations import SingleImputer, MultipleImputer, MiceImputer
si = SingleImputer() # pass through data once
mi = MultipleImputer() # pass through data multiple times
mice = MiceImputer() # pass through data multiple times and iteratively optimize imputations in each column

Which to use, and When?

There are tradeoffs between the three imputers
We won't get into the specifics regarding why, but here are a couple points to keep in mind:
Execution time (best to worst): SingleImputer, MultipleImputer, MiceImputer
Imputation quality (best to worst): MiceImputer, MultipleImputer, SingleImputer
This shouldn't come as a surprise. The MiceImputer does the most work, while the SingleImputer does the least
The example below use the MiceImputer, but you can swap in the MultipleImputer or SingleImputer as well

Imputations can be as simple as:

# simple example using default instance of MiceImputer
imp = MiceImputer()

# fit transform returns a generator by default, calculating each imputation method lazily
imp.fit_transform(data)

Or quite complex, such as:

# create a complex instance of the MiceImputer
# Here, we specify strategies by column and predictors for each column
# We also specify what additional arguments any `pmm` strategies should take
imp = MiceImputer(
    n=10,
    strategy={"salary": "pmm", "gender": "bayesian binary logistic", "age": "norm"},
    predictors={"salary": "all", "gender": ["salary", "education", "weight"]},
    imp_kwgs={"pmm": {"fill_value": "random"}},
    visit="left-to-right",
    return_list=True
)

# Because we set return_list=True, imputations are done all at once, not evaluated lazily.
# This will return M*N, where M is the number of imputations and N is the size of original dataframe.
imp.fit_transform(data)

Autoimpute also extends supervised machine learning methods from scikit-learn and statsmodels to apply them to multiply imputed datasets (using the MiceImputer under the hood). Right now, Autoimpute supports linear regression and binary logistic regression. Additional supervised methods are currently under development.

As with Imputers, Autoimpute's analysis methods can be simple or complex:

from autoimpute.analysis import MiLinearRegression

# By default, use statsmodels OLS and MiceImputer()
simple_lm = MiLinearRegression()

# fit the model on each multiply imputed dataset and pool parameters
simple_lm.fit(X_train, y_train)

# get summary of fit, which includes pooled parameters under Rubin's rules
# also provides diagnostics related to analysis after multiple imputation
simple_lm.summary()

# make predictions on a new dataset using pooled parameters
predictions = simple_lm.predict(X_test)

# Control both the regression used and the MiceImputer itself
mice_imputer_arguments = dict(
    n=3,
    strategy={"salary": "pmm", "gender": "bayesian binary logistic", "age": "norm"},
    predictors={"salary": "all", "gender": ["salary", "education", "weight"]},
    imp_kwgs={"pmm": {"fill_value": "random"}},
    visit="left-to-right"
)
complex_lm = MiLinearRegression(
    model_lib="sklearn", # use sklearn linear regression
    mi_kwgs=mice_imputer_arguments # control the multiple imputer
)

# fit the model on each multiply imputed dataset
complex_lm.fit(X_train, y_train)

# get su

Autoimpute

Install / Use

README

Autoimpute

Conference Talks

Notes on Development

Creators and Maintainers

Installation

Motivation

Main Features

Imputation Methods Supported

Todo

Example Usage

Which to use, and When?