Continuous missing values imputation using Autoencoders

Implementation of imputation using different types of autoencoders described in paper. This python package contains implementation of Denoising Autoencoder (autoencoder_impute.DaeImpute), Multimodal Autoencoder (autoencoder_impute.MaeImpute), Variational Autoencoder (autoencoder_impute.VaeImpute)

Installation

pip install autoencoder-impute

Usage

Denoising autoencoder (DAE)

Denoising Autoencoder (autoencoder_impute.DaeImpute)

from autoencoder_impute import DaeImpute

# Initialization
dae = DaeImpute(k=5, h=7)

# Training
dae.fit(df_train, df_val)

# Imputation
reconstructed_df = dae.transform(df_missing)

# Raw DAE output
raw_dae_out_df = dae.predict(df_missing)

# Save model
dae.save('dataset-k-5-h-5', path='.')

# Load model
new_dae = DAE(k=5, h=7)
new_dae.load('dataset-k-5-h-5', path='.')

Multimodal autoencoder (MAE)

Multimodal Autoencoder (autoencoder_impute.MaeImpute)

from autoencoder_impute import MaeImpute

# Initialization
mae = MaeImpute(k=5, h=7)

# Training
mae.fit(df_train, df_val)

# Imputation
reconstructed_df = mae.transform(df_missing)

# Raw DAE output
raw_dae_out_df = mae.predict(df_missing)

# Save model
mae.save('dataset-k-5-h-5', path='.')

# Load model
new_dae = MAE(k=5, h=7)
new_dae.load('dataset-k-5-h-5', path='.')

Variational autoencoder (VAE)

Variational Autoencoder (autoencoder_impute.VaeImpute)

from autoencoder_impute import VaeImpute

# Initialization
vae = VaeImpute(k=5, h=7)

# Training
vae.fit(df_train, df_val)

# Imputation
reconstructed_df = vae.transform(df_missing)

# Raw DAE output
raw_dae_out_df = vae.predict(df_missing)

# Save model
vae.save('dataset-k-5-h-5', path='.')

# Load model
new_dae = VAE(k=5, h=7)
new_dae.load('dataset-k-5-h-5', path='.')