sohamghosh121 / unsupervised-backpropagation

Matrix Imputation using Unsupervised Backpropagation

unsupervised-backpropagation

Matrix Imputation using Unsupervised Backpropagation

This is a Theano based Python implementation of the unsupervised backpropagation algorithm described in https://arxiv.org/pdf/1312.5394.pdf for matrix imputation.

unsupervised-backpropagation's Issues

Use with Categorical Data

Hello,

I am trying to use the UBP.py class with categorical data, and it is outputting all 1's. I am aware that the publication says to one-hot encode the categories and predict each one, selecting the mode from the categorical distribution. However, I am unsure how to do so with the UBP.py script.

I tried running it as-is, and I think that it is saving the output to X_reconstructed. Is that correct? If so, I tried to return X_reconstructed and all the values in the dataset were 1.0, even the values that did not have missing data.

Here is the code that I used to run it. I added this to the end of init() in the UBP.py class:

self.impute(self.X,  3, 3, [100, 100, 100])
imputed = self.predict()
print(imputed)
print(imputed.shape)

def predict(self):
    return self.X_reconstructed

Do you know how I could get it to work with categorical data, or why it would be outputting an array of all 1's?

Thanks,

-Bradley

RMSE: nan

Hello @sohamghosh121
I've tried to run your script but had trouble, here's my test example:

from UBP import Imputer
import pandas as pd
import numpy as np

X = pd.DataFrame([[1,2,3,4,5, 6, np.nan], [5,5,5,5,5, np.nan, 5]]).T.astype(np.float32).values

imp = Imputer(X, 1, 1, [10])
imp.impute()

Here's the result I get and it keeps looping forever

* Running imputation
* Initial RMSE: 3.55216297909
* #### Phase  1
Epoch took 6.776763281474511e-05 m
Epoch took 5.038633244112134e-05 m
Epoch took 4.993090018009146e-05 m
Epoch took 4.95495837337027e-05 m
Epoch took 4.9610216713820894e-05 m
Epoch took 5.0451583229005335e-05 m
Epoch took 4.677378262082736e-05 m
Epoch took 4.657153234196206e-05 m
Epoch took 4.653114980707566e-05 m
Epoch took 4.677001658516626e-05 m
* Epochs: 10 	 RMSE: nan

Do you have any comment on fixing this? Can you give me usage examples please?
Thanks